PJFP.com

Pursuit of Joy, Fulfillment, and Purpose

Tag: Terminal-Bench 2.1

  • Inkling: Thinking Machines Lab Releases Its First Open-Weights Model, a 975B Multimodal Mixture-of-Experts With Controllable Thinking Effort That Can Fine-Tune Itself on Tinker

    Thinking Machines Lab, the AI startup founded by former OpenAI CTO Mira Murati, has released Inkling, its first open-weights model trained from scratch. Inkling is a 975 billion parameter Mixture-of-Experts transformer (41B active) with a context window of up to 1 million tokens, native multimodal reasoning over text, images, and audio, and a dial for controllable thinking effort. The lab is explicit that Inkling is not the strongest model in the world. It is pitched as something arguably more useful: a broad, balanced, customizable foundation you can fine-tune on Tinker, with the full weights on Hugging Face. The announcement even includes a demo where Inkling fine-tunes itself and swaps in its own new weights.

    TLDR

    Thinking Machines Lab released Inkling, a 975B-total, 41B-active Mixture-of-Experts model pretrained on 45 trillion tokens of text, images, audio, and video, alongside a preview of Inkling-Small (276B total, 12B active). The release covers the model’s generalist benchmark profile across reasoning, agentic coding, tool use, vision, and audio; a controllable thinking effort setting that lets developers trade performance against tokens (matching Nemotron 3 Ultra on Terminal Bench 2.1 at roughly a third of the tokens); an encoder-free multimodal architecture using dMel spectrograms and hMLP image patches; a training recipe combining Muon and Adam with weight decay coupled to the learning rate; RL scaled past 30 million rollouts with log-linearly improving reasoning and an emergent compression of the chain of thought; an epistemics push covering calibration, forecasting (where it beats several frontier models), abstention, and censorship resistance; the strongest FORTRESS adversarial safety score among compared open-weights models; a headline-grabbing demo of the model fine-tuning itself into a lipogram assistant via Tinker; and day-one availability on Tinker (at a 50% discount), Hugging Face, and inference partners including Together, Fireworks, Modal, Databricks, Baseten, vLLM, SGLang, and llama.cpp.

    Thoughts

    The most striking thing about this launch is its honesty. Nearly every frontier release leads with a claim to be the best at something, and the fine print walks it back. Thinking Machines Lab says plainly that Inkling is not the strongest model available, open or closed, and then makes the case that “strongest” is the wrong axis for most real buyers. If you are going to run a model millions of times inside a product, what you care about is the cost curve, the adaptability, and whether you can shape it to your workflow. That framing conveniently matches their business (Tinker sells fine-tuning), but it also matches how production AI actually gets deployed, where cost and latency are binding constraints and a benchmark crown is trivia.

    The self-fine-tuning demo deserves more attention than it will probably get. Asked to become a lipogram assistant that never uses the letter “e” (a behavior prompting alone cannot reliably produce), Inkling wrote its own training objective and scoring function, generated its own synthetic data, launched the run on Tinker, evaluated the result against its base self, and then staged a weight swap so the improved checkpoint took over the session. That is a closed loop of specify, train, evaluate, and self-update, packaged as a cute product demo. The loop is the primitive behind every serious conversation about recursive self-improvement, and here it is running as a marketing asset with a 27 minute wall clock. The gap between “toy objective” and “economically meaningful objective” is now a question of reward design, not plumbing.

    Controllable thinking effort is the feature I expect developers to care about most. Instead of publishing a single score, TML publishes a curve: sweep the effort setting from 0.2 to 0.99 and watch performance trade against generated tokens. Inkling reportedly matches Nemotron 3 Ultra on Terminal Bench 2.1 while spending about a third of the tokens. Benchmarks reported as single points hide exactly this, and a model that reaches a target score cheaply beats a model that scores two points higher at triple the cost in any high-volume workload. Expect effort curves to become standard marketing for open models, the way context length became standard a couple of years ago.

    The epistemics section is quietly the most differentiated part of the release. TML trained calibration directly, running RL against proper scoring rules on resolved real-world questions, and pairing a rubric grader with a claims grader that does agentic web search to verify each factual assertion. The result is a model that beats GPT-5.5 and Claude Opus 4.8 on ForecastBench without search and holds its own on Prophet Arena. A model that knows when to say “I don’t know” is more useful across messy real-world domains than one that confabulates confidently, and it is notable that a lab whose stated mission is extending human will and judgment treats calibrated uncertainty as a first-class training target rather than a safety afterthought. The censorship-resistance training, validated on Cognition’s Propaganda and Censorship Eval, extends the same idea: trustworthiness as a capability you train, not a policy you bolt on.

    Finally, the open-weights safety tension is handled with unusual candor. Inkling posts the strongest adversarial FORTRESS score among the open models compared while keeping benign over-refusal low, and it was tested externally for CBRN, cyber, and loss-of-control capabilities. But everyone in this space knows fine-tuning can strip safety behavior from open weights, and TML ships a fine-tuning platform for this exact model. Their acknowledgment that they are actively studying how safety behavior survives fine-tuning on Tinker is the right thing to say, and it is also the open question that will define whether “safe open weights” is a coherent category at all.

    Key Takeaways

    • Inkling is Thinking Machines Lab’s first from-scratch, open-weights model: a Mixture-of-Experts transformer with 975B total parameters, 41B active, and a context window up to 1M tokens.
    • It was pretrained on 45 trillion tokens spanning text, images, audio, and video, and reasons natively over text, images, and audio without separate encoders.
    • A preview of Inkling-Small ships alongside it: a 276B-parameter MoE with just 12B active parameters that matches or beats its larger sibling on several benchmarks thanks to an improved pretraining recipe.
    • TML explicitly positions Inkling as a base for customization rather than the strongest overall model, leaning on multimodality, efficient thinking, and Tinker fine-tuning as the differentiators.
    • The launch demo shows Inkling fine-tuning itself: it wrote its own training objective and data, ran the job through the Tinker API, evaluated the result, and hot-swapped to its own new weights inside the OpenCode harness.
    • The self-fine-tuning target was a lipogram assistant that never uses the letter “e,” a behavior chosen precisely because prompting alone cannot reliably achieve it; the full loop completed in about 27 minutes.
    • Controllable thinking effort is a core feature: a setting swept from 0.2 to 0.99 traces a full performance-versus-tokens curve instead of a single benchmark point.
    • On Terminal Bench 2.1, Inkling matches Nemotron 3 Ultra’s score at roughly one third of the generated tokens, the release’s flagship efficiency claim.
    • Inkling was trained to run inside a variety of coding and agent harnesses, with tool sets and schemas randomized during training to reduce sensitivity to any particular harness.
    • On Design Arena’s blinded human-evaluated Agentic Web Dev leaderboard, Inkling scores 1257, among the strongest open-weights models and tied with Claude Opus 4.6.
    • Headline benchmark scores at effort 0.99 include SWEBench Verified 77.6%, SWEBench Pro Public 54.3%, Terminal Bench 2.1 63.8%, GPQA Diamond 87.2%, AIME 2026 97.1%, and HLE 29.7% text-only (46.0% with tools).
    • Agentic and general scores include MCP Atlas 74.1%, Tau 3 Banking 23.7%, and BrowseComp 77.1% with context management.
    • Vision results are strong for an open model: MMMU Pro 73.5%, CharXiv RQ 78.1%, rising to 82.0% when the model uses a Python tool for zooming and cropping during visual reasoning.
    • Audio results place it among the strongest open-weights audio models: VoiceBench 91.4%, MMAU 77.2%, and Audio MC 56.6%, well ahead of Qwen3-Omni and Nemotron Nano-Omni on the last.
    • The multimodal stack is encoder-free: audio enters as discrete dMel spectrograms and images as 40×40 pixel patches through a four-layer hMLP, both passed through a lightweight embedding layer and processed jointly with text tokens.
    • The MoE design largely follows DeepSeek-V3: 256 routed experts plus 2 shared experts per layer, 6 routed experts active per token, with a sigmoid router and auxiliary-loss-free load balancing.
    • Attention interleaves sliding-window and global layers at a 5:1 ratio with 8 KV heads, and uses a learned relative positional embedding instead of RoPE, which TML found extrapolates better to long sequences.
    • Short convolutions are applied after the key and value projections and on the attention and MLP residual branch outputs, an unusual architectural touch aimed at efficiency and long-context performance.
    • Training used a hybrid optimizer strategy, Muon for large matrix weights and Adam for everything else, with weight decay coupled to the square of the learning rate to keep weight magnitudes stable.
    • Post-training was bootstrapped with a small SFT phase on synthetic data generated by open-weights models including Kimi K2.5, with the large majority of compute spent on large-scale RL.
    • RL was scaled past 30 million rollouts across two long continuous runs, with reasoning performance on a held-out aggregate (AIME, HLE, GPQA, and others) improving log-linearly throughout.
    • Effort control was trained by varying the system message and per-token cost across rollouts, teaching the model to modulate its own thinking budget.
    • An emergent effect appeared during RL: the chain of thought compressed over training, dropping articles and connectives into a telegraphic style, driven purely by efficiency pressure rather than any targeted reward.
    • Inkling was TML’s first major training effort and ran on NVIDIA GB300 NVL72 systems; the lab says future models will push compute scale further across pretraining and RL.
    • Calibration was trained directly with RL against proper scoring rules on a large corpus of resolved real-world questions, treating well-placed confidence as a capability rather than a byproduct.
    • On ForecastBench without search, Inkling’s Brier Index of 61.1 beats GPT-5.5 (59.1) and Claude Opus 4.8 (54.6), and it stays competitive with search enabled and on Prophet Arena.
    • Instruction following was trained with two automated graders working together: a rubric grader scoring against a checklist and a claims grader that verifies each factual claim via agentic web search, improving helpfulness and reducing hallucination simultaneously.
    • Abstention-aware rewards on short-form factual QA taught the model to answer when confident and hedge or decline when not, with some prompts explicitly forcing or forbidding hedging so the user’s preference wins.
    • Inkling was trained to answer directly on topics subject to censorship, and Cognition’s Propaganda and Censorship Eval found strong censorship non-compliance.
    • On FORTRESS, Inkling posts the strongest adversarial refusal score (78.0%) of any compared open-weights model while keeping benign compliance high (95.9%), and scores 98.6% on StrongREJECT.
    • Safety testing covered CBRN, cyber, and loss-of-control capabilities plus human-AI threat vectors like sycophancy, vulnerable users, and manipulation, verified by commissioned external testers.
    • Inkling is available for fine-tuning on Tinker today with 64K and 256K context options at a 50% limited-time discount, plus a free Inkling Playground chat interface in the Tinker console.
    • Full weights are on Hugging Face, including an NVFP4 checkpoint for efficient inference on NVIDIA Blackwell, with API availability via Together, Fireworks, Modal, Databricks, and Baseten and inference support in SGLang, vLLM, TokenSpeed, and llama.cpp.
    • TML frames Inkling as the first in a family and as the intended background reasoning model for its previously announced real-time interaction models system.

    Detailed Summary

    What Inkling Is and Why It Exists

    Thinking Machines Lab frames its mission as building AI that extends human will and judgment, and Inkling as the logical next step after shipping the Tinker customization platform, previewing an interaction-focused AI system, and publishing research. Inkling is a Mixture-of-Experts transformer with 975B total and 41B active parameters, a context window up to 1M tokens, and pretraining on 45 trillion tokens of mixed text, image, audio, and video data. The lab is upfront that it is not the strongest model available. The pitch is breadth plus adaptability: a generalist trained across agentic, reasoning, coding, instruction-following, factuality, vision, and audio tasks rather than tuned to dominate one leaderboard, offered with full weights so people can make it their own. It launches with a preview sibling, Inkling-Small, at 276B total and 12B active parameters.

    The Self-Fine-Tuning Demo

    To demonstrate what customization means, TML asked Inkling to fine-tune itself. Running inside the OpenCode harness with access to Tinker, the model was told to become a lipogram assistant that never uses the letter “e.” Inkling drafted the plan, wrote an objective file with a scoring function (any response containing “e” scores zero), generated synthetic training data, launched a supervised fine-tuning run through the Tinker API, evaluated the checkpoint against its base self, and then staged a self-update so the supervisor relaunched the session on the new weights. The pipeline passed in about 27 minutes, and the updated model answered a test question about launching an LLM without a single “e.” It is a whimsical objective wrapped around a serious primitive: a model autonomously specifying, running, and adopting its own weight updates.

    Agentic Coding and Tool Use

    TML trained Inkling to operate inside many coding and agent harnesses, randomizing tool sets and schemas during training so the model does not overfit to one environment. The release showcases three demos: a one-shot job-application web app that then hosts an embedded browser-use agent operating its own interface; a nine-page, cohesively designed PDF food and travel journal produced from a single editorial prompt with web-verified details; and a server-authoritative multiplayer snake game refined over 40 iterations of feedback from GPT Codex acting as a reviewer. On benchmarks, Inkling posts 77.6% on SWEBench Verified, 54.3% on SWEBench Pro Public, and 63.8% on Terminal Bench 2.1, competitive within the open-weights field, and 1257 on Design Arena’s human-judged web dev leaderboard, in the same band as Claude Opus 4.6.

    Controllable Thinking Effort

    Rather than reporting a single operating point, TML sweeps Inkling’s effort setting from 0.2 to 0.99 and plots score against mean generated tokens on Terminal Bench 2.1, HLE, and IFBench, with competitors shown at their default settings. The headline result is efficiency: Inkling reaches Nemotron 3 Ultra’s Terminal Bench score at roughly a third of the tokens. The argument is that cost and latency are binding constraints in production, especially for interactive collaboration, so the full cost curve, not the peak score, is what developers should evaluate. Effort can be set from within the agent harness, and the ability was trained by varying system messages and per-token costs across RL rollouts.

    Native Multimodality Without Encoders

    Inkling is designed to serve as the background reasoning model for TML’s interaction models system, which requires real-time voice and vision collaboration. The multimodal components are trained from scratch with an encoder-free architecture: audio arrives as discrete dMel spectrograms and images as 40×40 pixel patches through a four-layer hMLP, both mapped through a lightweight embedding layer and processed jointly with text. The model transcribes speech, follows spoken instructions, reasons over long recordings, and answers questions about charts and diagrams, optionally using a Python tool to zoom and crop images mid-reasoning. Scores like 91.4% on VoiceBench and 82.0% on CharXiv RQ with Python place it among the strongest open-weights multimodal models, though still behind Gemini 3.1 Pro.

    Epistemics: Calibration, Forecasting, and Censorship Resistance

    TML groups calibration, instruction following, and censorship resistance under the banner of epistemics. Calibration was trained with RL against proper scoring rules on resolved real-world questions, and it shows: Inkling’s ForecastBench Brier Index of 61.1 without search beats GPT-5.5 and Claude Opus 4.8, and its Prophet Arena score sits close to the frontier. Instruction following used two complementary automated graders, a rubric checklist and a claims grader that verifies factual assertions through agentic web search, so recall-spraying to hack rubrics gets penalized by the factuality check. Targeted abstention-aware QA datasets taught the model to say “I don’t know” or give hedged best guesses when appropriate, while still complying when a user demands a forced guess. Finally, the model was trained to answer directly on censorship-prone topics, with Cognition’s Propaganda and Censorship Eval finding strong non-compliance with censorship patterns.

    Safety for an Open-Weights Release

    Inkling was trained to an internal behavioral spec across all modalities and then checked by commissioned external safety testers. Evaluations covered dangerous capabilities (CBRN, cyber, loss of control) and human-AI threat vectors including sycophancy, vulnerable users, and harmful manipulation. On FORTRESS, which pairs adversarial harmful requests with benign look-alikes, Inkling posts the strongest adversarial score among the compared open models (78.0%) without collapsing on the benign side (95.9%), and it scores 98.6% on StrongREJECT. TML acknowledges the open question hanging over every open-weights release: how safety behavior holds up under fine-tuning, which it says it is actively studying on Tinker.

    Architecture and Training Recipe

    The MoE layout follows DeepSeek-V3: 256 routed experts and 2 shared experts per layer with 6 routed experts active per token, a sigmoid-based router, and auxiliary-loss-free load balancing. Attention interleaves sliding-window and global layers 5:1 with 8 KV heads, and positions are encoded with a learned relative positional embedding that TML found outperforms and out-extrapolates RoPE. Short convolutions appear after the key and value projections and on residual branch outputs. Optimization was hybrid, Muon for large matrices and Adam elsewhere, with hyperparameter schedules drawn from the lab’s modular manifolds research and weight decay coupled to the square of the learning rate to keep weight norms stable. Post-training bootstrapped from a small SFT phase on synthetic data from open models including Kimi K2.5, then spent the bulk of compute on large-scale RL. Everything ran on NVIDIA GB300 NVL72 systems.

    RL at Scale and the Emergent Compression of Thought

    TML scaled asynchronous RL past 30 million rollouts across two long continuous runs, with performance on a held-out aggregate of reasoning evals improving log-linearly the whole way. Along the way an unplanned behavior emerged: the chain of thought became progressively more concise, shedding grammatical overhead into a telegraphic style (“We need to understand” becomes “We need determine”) while remaining comprehensible and leaving final answers unaffected. No reward targeted this; token efficiency pressure alone drove the compression, echoing an observation Cognition made while training SWE-1.7. It is a vivid example of optimization discovering its own shorthand.

    Inkling-Small

    The preview of Inkling-Small is arguably the sleeper story: with 12B active parameters against Inkling’s 41B, it matches or exceeds the larger model on a surprising number of benchmarks, including GPQA Diamond (88.3% vs 87.2%), IFBench (83.4% vs 79.8%), and CharXiv RQ with Python (83.4% vs 82.0%). TML attributes this to pretraining data and recipe improvements made after the big model trained, with both models sharing the same post-training stack. The clearest gaps favoring big Inkling are factuality (SimpleQA 43.9% vs 20.9%), Terminal Bench, and Tau 3 Banking. Full weights for Inkling-Small will be released once testing finishes, and its cost and latency profile targets high-volume workloads like coding, LLM grading, and synthetic data generation.

    Availability and the Ecosystem Play

    Inkling is on Tinker today with 64K and 256K context options at a limited-time 50% discount, plus a free Inkling Playground chat interface with integrated web search in the Tinker console so developers can get a feel for the model before committing to a run. The cookbook gained native Inkling support and three new audio recipes, and a new tml-renderer handles chat templates, tool calls, reasoning content, and multimodal inputs. Deployment partnerships span Together, Fireworks, Modal, Databricks, and Baseten for APIs; RadixArk for SGLang and Miles; Inferact for vLLM; Lightseek for TokenSpeed; Unsloth for llama.cpp; and Hugging Face for transformers integration. Full weights are on Hugging Face in both the original checkpoint and an NVFP4 checkpoint for NVIDIA Blackwell inference.

    Notable Quotes

    “Our mission is to build AI that extends human will and judgment.”

    Thinking Machines Lab, opening the Inkling announcement

    The company’s north star, and the lens through which the whole release (customization, calibration, open weights) is framed.

    “Inkling is not the strongest overall model available today, open or closed. Instead, a combination of qualities makes it a good open-weights base for customization: multimodal capabilities, efficient thinking, and availability on Tinker for fine-tuning.”

    Thinking Machines Lab, positioning the release

    A rare piece of launch-day honesty from a frontier lab, and the strategic thesis of the whole release.

    “Picking the right base model to fine-tune is a qualitative judgment that combines measurable benchmarks with the unique feel of a model that comes from playing with it.”

    Thinking Machines Lab, on why the Inkling Playground exists

    An argument that vibes are data, from the lab that built a playground into a fine-tuning console.

    “Cost and latency are often binding constraints in real-world applications, and low latency in particular is crucial for enabling collaboration and improvement through iteration.”

    Thinking Machines Lab, on controllable thinking effort

    The case for evaluating models on their full effort-versus-performance curve instead of a single benchmark point.

    “A model that’s confident in every answer it gives, including when it’s missing info and confabulates, forces the user to double-check everything.”

    Thinking Machines Lab, on why calibration was a training target

    The clearest one-line justification for treating calibrated uncertainty as a capability rather than a nicety.

    “Together, the two graders improve helpfulness and reduce hallucination at the same time, rather than trading one for the other.”

    Thinking Machines Lab, on pairing a rubric grader with a web-searching claims grader

    A neat solution to rubric hacking: verify every claim with agentic search so spraying plausible facts stops paying.

    “Safety is crucial for open-weights models. We’re continuing to study safety behavior and capability uplift in customizable models, including how safety behavior is impacted by fine-tuning on Tinker.”

    Thinking Machines Lab, on the open question of fine-tunable safety

    The acknowledgment that safety trained into open weights must survive the very customization the product sells.

    “Inkling is just the start: our first release in a model family we will continue to build on.”

    Thinking Machines Lab, on the roadmap

    Together with the GB300 compute note, a clear signal that larger and stronger family members are coming.

    Read the full announcement, including the interactive demos, effort curves, and complete benchmark tables, on the Thinking Machines Lab blog.

    Related Reading

  • Claude Opus 4.8 Released: Anthropic Bets on Honesty, Dynamic Workflows, Effort Control, and Cheaper Fast Mode

    Anthropic has released Claude Opus 4.8, the newest member of its flagship Opus class, available today across every surface and priced exactly like the model it replaces. The company calls it “a modest but tangible improvement” on Opus 4.7, but the framing undersells what is actually interesting here: the headline upgrade is not a benchmark number, it is honesty. Opus 4.8 is built to know when it does not know, and that single behavioral shift may matter more for real agent work than any raw capability bump.

    TLDR

    Claude Opus 4.8 is an across-the-board upgrade to Anthropic’s Opus class that ships today at the same regular price as Opus 4.7 ($5 per million input tokens, $25 per million output tokens), with the model positioned as “a more effective collaborator.” The marquee improvement is honesty: Opus 4.8 is roughly four times less likely than its predecessor to let flaws in its own code pass unremarked, and it is more willing to flag uncertainty rather than confidently claim progress on thin evidence. A pre-release alignment assessment found new highs on prosocial traits like supporting user autonomy and acting in the user’s best interest, with misaligned behavior at rates similar to Anthropic’s best-aligned model, Claude Mythos Preview. Three things launch alongside the model: dynamic workflows in Claude Code (research preview), where Claude plans work then runs hundreds of parallel subagents that run even longer and verify their own outputs before reporting back; effort control in claude.ai and Cowork, a slider for how hard Claude thinks; and a Messages API update that accepts system entries inside the messages array so developers can update instructions mid-task without breaking the prompt cache. Fast mode now runs at 2.5x speed and is three times cheaper than before ($10 / $50 per million tokens). The roadmap points to cheaper Opus-equivalent models, a higher-intelligence class above Opus, and a wider rollout of Mythos-class models gated behind stronger cyber safeguards under Project Glasswing.

    Thoughts

    The most important sentence in this announcement is not about coding scores. It is the claim that Opus 4.8 is about four times less likely than Opus 4.7 to let flaws in its own code slip by without comment. For a chat assistant, overconfidence is annoying. For an agent, it is catastrophic. The whole premise of long-running autonomous work is that you hand the model a task and walk away, which means the model’s own judgment about whether it succeeded becomes the only judgment in the loop until you come back. A model that confidently declares victory on a half-finished migration does not save you time, it costs you a debugging session plus the time you spent trusting it. Honesty, framed this way, is not a soft virtue. It is the load-bearing reliability property that makes unattended agents usable at all.

    Read the launch as a single coherent argument rather than a list of features, and the pieces lock together. Dynamic workflows let Claude plan a job and fan out hundreds of parallel subagents that, with Opus 4.8, run longer than before. Effort control lets you dial up how much the model thinks. The honesty improvement means the model checks its own work and flags what it is unsure about instead of papering over it. Put those three together and you get one product thesis: let it run longer, let it think harder, and trust it to tell you when something is wrong. The codebase-scale migration example, hundreds of thousands of lines from kickoff to merge with the existing test suite as the bar, is the proof point. None of those three capabilities is worth much alone. A model that runs for hours but lies about its results is a liability. A model that flags uncertainty but cannot sustain a long task never reaches the moment where its honesty matters. Anthropic shipped all three at once because they only pay off together.

    The economics deserve a closer look than the “same price” headline invites. Regular pricing is flat versus Opus 4.7, which is the polite way of saying you get a better model for free. The real move is fast mode: 2.5x the speed at three times cheaper than it cost on previous models, landing at $10 per million input and $50 per million output. That is Anthropic quietly attacking the latency-versus-cost tradeoff that has shaped how teams deploy frontier models. Until now, “fast” meant “expensive,” so you reserved it for interactive moments and ate the wait everywhere else. Collapsing that premium changes the default. And note the subtle token story underneath: Opus 4.8 at its default high effort spends roughly the same tokens on coding as Opus 4.7’s default while performing better, so the effort slider is not a way to bleed you dry, it is an honest exposure of the quality-cost dial that was always there implicitly.

    The Messages API change is the kind of unglamorous plumbing that practitioners will appreciate immediately. Letting system entries live inside the messages array means you can update an agent’s instructions, permissions, token budget, or environment context partway through a task without smuggling the update through a fake user turn and without blowing up your prompt cache. Anyone who has built a long-running agent has hit this wall: the world changes mid-task, the agent needs new constraints, and the only clean way to inject them previously was a cache-busting hack. This is Anthropic treating agents as first-class, stateful, long-lived processes rather than oversized chat sessions. It is a small spec change with outsized implications for how you architect an agent that runs for an hour.

    Then there is the roadmap, where the most telling line is the quietest. Anthropic says a small number of organizations are already using Claude Mythos Preview for cybersecurity work under Project Glasswing, and that models of this capability level require stronger cyber safeguards before general release. Notice that they are pinning Opus 4.8’s alignment numbers to Mythos as the benchmark for “best-aligned,” while simultaneously holding Mythos back from general availability on safety grounds. That is a deliberate signal: the next class of model is good enough that they are gating it on cyber-offense risk, not on capability. For a site about the pursuit of joy, fulfillment, and purpose through AI, this is the part worth sitting with. The frontier is increasingly defined not by what the models can do, but by what their builders decide it is responsible to ship. Honesty in the small (flagging a bad line of code) and restraint in the large (holding back a cyber-capable model) are the same instinct expressed at two different scales.

    Key Takeaways

    • Claude Opus 4.8 is now available everywhere, replacing Opus 4.7 as Anthropic’s flagship Opus-class model and positioned as “a more effective collaborator.”
    • Regular usage pricing is unchanged from Opus 4.7, holding at $5 per million input tokens and $25 per million output tokens, so the capability gains come at no added cost.
    • The single most emphasized improvement is honesty, which Anthropic treats as a core trained behavior rather than a marketing flourish.
    • Evaluations show Opus 4.8 is around four times less likely than its predecessor to let flaws in its own code pass unremarked, a direct reliability win for autonomous coding.
    • Early testers report the model is more likely to flag uncertainty about its work and less likely to make unsupported claims or jump to conclusions on thin evidence.
    • A detailed alignment assessment was run before release and concluded Opus 4.8 reaches new highs on prosocial traits like supporting user autonomy and acting in the user’s best interest.
    • Misaligned behavior such as deception or cooperation with misuse is at rates substantially lower than Opus 4.7 and similar to Anthropic’s best-aligned model, Claude Mythos Preview.
    • The full alignment assessment and pre-deployment safety tests are documented in the public Claude Opus 4.8 System Card.
    • Dynamic workflows launch as a research preview inside Claude Code, letting Claude plan the work and then run hundreds of parallel subagents in a single session.
    • With Opus 4.8, those subagents can run even longer, and Claude verifies its outputs before reporting back rather than declaring success blindly.
    • Anthropic’s flagship example for dynamic workflows is a codebase-scale migration across hundreds of thousands of lines of code, from kickoff to merge, using the existing test suite as the success bar.
    • Dynamic workflows are available in Claude Code for the Enterprise, Team, and Max plans.
    • Effort control arrives in claude.ai and Cowork as a setting next to the model selector that lets users choose how much effort Claude puts into a response.
    • Higher effort makes Claude think more frequently and deeply for better answers; lower effort responds faster and consumes rate limits more slowly. Effort control is available on all plans.
    • Opus 4.8 defaults to “high” effort, judged the best overall balance of quality and user experience.
    • On coding tasks, the default effort spends a similar number of tokens as Opus 4.7’s default but delivers better performance, so quality rises without a token penalty.
    • Users can select “extra” (called “xhigh” in Claude Code) or “max” to spend more tokens for stronger results, and Anthropic recommends “extra” for difficult tasks and long-running asynchronous workflows.
    • Rate limits in Claude Code were increased to accommodate the higher token usage of the higher effort levels.
    • The Messages API now accepts system entries inside the messages array, a meaningful change for agent developers.
    • That update lets developers change Claude’s instructions mid-task, adjusting permissions, token budgets, or environment context, without breaking the prompt cache or routing through a user turn.
    • Fast mode now runs at 2.5x speed and is three times cheaper than it was for previous models, priced at $10 per million input tokens and $50 per million output tokens.
    • Developers access the model as claude-opus-4-8 through the Claude API.
    • Partner Miguel Gonzalez reports Opus 4.8 scored 84% on Online-Mind2Web, a meaningful jump over both Opus 4.7 and GPT-5.5, calling it the strongest computer-use and browser-agent model his team has tested.
    • Databricks reports that, inside Genie, Opus 4.8 reasons over unstructured content like PDFs and diagrams at 61% cheaper token cost than Opus 4.7.
    • Thomson Reuters reports Opus 4.8 is the first model to break 10% overall on the all-pass standard of its Legal Agent Benchmark, the highest score recorded there.
    • Eleven partners weighed in, including Cursor, Cognition’s Devin, Databricks Genie, Thomson Reuters CoCounsel, and Hebbia, spanning coding, legal, finance, and enterprise data work.
    • Anthropic is working on models that deliver many of the same capabilities as Opus at a lower cost.
    • The company plans to release a new class of model with even higher intelligence than Opus.
    • Under Project Glasswing, a small number of organizations are already using Claude Mythos Preview for cybersecurity work, with Mythos-class models expected to reach all customers in the coming weeks once stronger cyber safeguards are in place.

    Detailed Summary

    What Claude Opus 4.8 Is

    Claude Opus 4.8 is an upgrade to Anthropic’s Opus class of models, building on Opus 4.7 with improvements across benchmarks covering coding, agentic skills, reasoning, and practical knowledge-work tasks. Anthropic describes the result as “a more effective collaborator” while characterizing the release overall as “a modest but tangible improvement on its predecessor.” The model is available today, everywhere, and developers call it as claude-opus-4-8 via the Claude API. The announcement includes a comparison table against the predecessor and other models, though the per-cell numbers in that table are published as an image and are not reproduced here as text.

    Honesty: The Headline Improvement

    Anthropic singles out honesty as one of the most prominent improvements in Opus 4.8. All of the company’s models are trained to be honest, which includes avoiding claims they cannot support. A persistent problem with AI models generally is that they sometimes jump to conclusions, confidently claiming progress despite thin evidence. Early testers report that Opus 4.8 is more likely to flag uncertainties about its own work and less likely to make unsupported claims. The most concrete measure: evaluations show Opus 4.8 is around four times less likely than its predecessor to allow flaws in code it has written to pass unremarked. For agentic and unattended use, this self-skepticism is the difference between a model that reliably tells you when something went wrong and one that quietly ships a broken result.

    Alignment Assessment

    A detailed alignment assessment was run before release. On the positive side, the Alignment team concluded that Opus 4.8 “reaches new highs on our measures of prosocial traits like supporting user autonomy and acting in the user’s best interest.” On the risk side, misaligned behavior such as deception or cooperation with misuse occurs at rates substantially lower than Opus 4.7, and similar to Anthropic’s best-aligned model, Claude Mythos Preview. The full alignment assessment and the pre-deployment safety tests are published in the Claude Opus 4.8 System Card, which also contains the complete benchmark table and wider evaluations.

    Dynamic Workflows in Claude Code

    Launching today as a research preview in Claude Code, dynamic workflows let Claude plan the work and then run hundreds of parallel subagents in a single session. With Opus 4.8, those agents can run even longer than before, and Claude verifies its outputs before reporting back rather than reporting unchecked results. The showcase example is a codebase-scale migration: Claude Code with Opus 4.8 can carry out migrations across hundreds of thousands of lines of code, all the way from kickoff to merge, using the existing test suite as its bar for success. Dynamic workflows are available in Claude Code for the Enterprise, Team, and Max plans.

    Effort Control

    Effort control arrives in claude.ai and Cowork as a setting alongside the model selector that lets users choose how much effort Claude puts into a response. Higher effort means Claude thinks more frequently and deeply for better responses; lower effort means it responds faster and uses rate limits more slowly. Opus 4.8 defaults to “high” effort, which Anthropic judged the best overall balance of quality and user experience. On coding tasks, that default spends a similar number of tokens as Opus 4.7’s default while performing better. Users who want more can choose “extra” (called “xhigh” in Claude Code) or “max” to spend more tokens for stronger results, and Anthropic recommends “extra” for difficult tasks and long-running asynchronous workflows. To support the heavier token usage at higher effort levels, rate limits in Claude Code were increased. Effort control is available on all plans.

    Messages API Update

    The Messages API now accepts system entries inside the messages array. This lets developers update Claude’s instructions mid-task without breaking the prompt cache and without routing the update through a user turn. In practice that means you can update permissions, token budgets, or environment context while an agent is running, which is exactly the kind of statefulness a long-running autonomous process needs. It is a small specification change with significant consequences for how developers build durable agents.

    Pricing and Fast Mode

    Regular usage pricing is unchanged from Opus 4.7: $5 per million input tokens and $25 per million output tokens. The notable shift is in fast mode, where the model works at 2.5x the speed and fast mode is now three times cheaper than it was for previous models, landing at $10 per million input tokens and $50 per million output tokens. The combination of unchanged regular pricing and dramatically cheaper fast mode reshapes the latency-versus-cost calculus that has long governed how teams deploy frontier models.

    Partner Results Across Coding, Legal, Finance, and Data

    Eleven partners shared results spanning the spectrum of professional work. Miguel Gonzalez reports 84% on Online-Mind2Web, a meaningful jump over both Opus 4.7 and GPT-5.5, calling it the strongest computer-use and browser-agent model his team has tested. Databricks reports that Genie reasons over unstructured content like PDFs and diagrams at 61% cheaper token cost than Opus 4.7. Thomson Reuters reports Opus 4.8 is the first model to break 10% overall on the all-pass standard of its Legal Agent Benchmark. Cursor reports gains across every effort level on CursorBench with more efficient tool calling, and Cognition reports that Devin sees cleaner tool use, fixes to the comment-verbosity and tool-calling issues seen with Opus 4.7, and improvements over Opus 4.6. Hebbia reports strong quality with better citation precision and more token efficiency on retrieval for dense financial filings. The footnotes note that Terminal-Bench 2.1 was scored on the Terminus-2 public harness (GPT-5.5’s Codex CLI harness score is 83.4%), that OSWorld-Verified methodology changed with Opus 4.7’s score updated to 82.3%, and that on Finance Agent v2 Gemini 3.5 Flash scores 57.9%.

    What Is Next: Cheaper Models, Higher Intelligence, and Mythos

    Anthropic outlined a three-part roadmap. First, the company is working on models that provide many of the same capabilities as Opus at a lower cost. Second, it plans to release a new class of model with even higher intelligence than Opus. Third, as part of Project Glasswing, a small number of organizations are currently using Claude Mythos Preview for cybersecurity work; models of this capability level require stronger cyber safeguards before general release, and Anthropic expects to bring Mythos-class models to all customers in the coming weeks.

    Notable Quotes

    “Claude Opus 4.8 has noticeably better judgment. In Claude Code, it asks the right questions, catches its own mistakes, pushes back when a plan isn’t sound, and builds up confidence around complex, multi-service explorations before making big changes. It’s a great model to build with.”

    Tom Pritchard, Staff Engineer, in Claude Code

    “On our Super-Agent benchmark, Claude Opus 4.8 is the only model to complete every case end-to-end, beating prior Opus models and GPT-5.5 at parity on cost. For agent products in translation, deep research, slide-building, and analysis, it delivers powerful reliability.”

    Kay Zhu, Co-Founder and CTO, on the Super-Agent benchmark

    “On CursorBench, Claude Opus 4.8 exceeds prior Opus models across every effort level. Tool calling is meaningfully more efficient, using fewer steps for the same intelligence, and it carries end-to-end tasks through.”

    Michael Truell, Co-Founder and CEO, on CursorBench results

    “Claude Opus 4.8 delivers the highest score recorded on our Legal Agent Benchmark, and is the first model to break 10% overall on the all-pass standard. For substantive legal work, that’s the kind of accuracy lift that translates directly into how much real attorney work our customers can hand off with confidence.”

    Niko Grupen, Head of Applied Research, on the Legal Agent Benchmark

    “Claude Opus 4.8 feels like a major quality-of-life update over Opus 4.7: faster, easier to collaborate with, and better at carrying context and style direction across a long session. Opus 4.8 is the model I kept trusting for work where voice, taste, and technical execution all have to happen side-by-side.”

    Katie Parrott, Staff Writer, on long writing sessions

    “Claude Opus 4.8 is the strongest computer-use and browser-agent model we’ve tested, scoring 84% on Online-Mind2Web, which is a meaningful jump over both Opus 4.7 and GPT-5.5. It stays reflective and on-task in the way our customers’ agent workloads need to be reliable end-to-end.”

    Miguel Gonzalez, Tech Lead, on computer-use and browser agents

    “Claude Opus 4.8 uses tools cleanly and follows instructions with the consistency our autonomous engineering workloads need to keep running unattended. It improves on Opus 4.6 and fixes the comment-verbosity and tool-calling issues we saw with Opus 4.7. This release from Anthropic translates directly into faster capability gains for engineers building on Devin.”

    Scott Wu, CEO, on building with Devin

    “On our long-running evals, Claude Opus 4.8’s analysis was consistently higher quality than prior Opus models. It finished faster and produced richer, more information dense outputs. Overall, a noticeably better signal to noise ratio. The biggest differentiator was Opus 4.8’s tendency to proactively flag issues with the inputs and outputs of an analysis, something other models routinely missed and left to the users to catch.”

    Michael Ran, Sr. Investment Associate, on long-running analysis evals

    Claude Opus 4.8 is a quieter release than its “modest but tangible” billing suggests, because the gains land where autonomous work actually lives: a model that flags its own uncertainty, runs longer and checks itself, scales effort on demand, and stays affordable while fast mode gets cheaper. The honesty improvement alone changes the trust math for anyone deploying agents. Read Anthropic’s full announcement here.

    Related Reading