Anthropic Releases Claude Opus 4.8 With Sharper Judgment and Cheaper Fast Mode

Anthropic released Opus 4.8 on Thursday, the newest version of its most advanced publicly available model, arriving just 41 days after Opus 4.7. That's a compressed cycle by any measure, and context matters here: the fast turnaround may partly reflect a chilly reception to Opus 4.7, while the interval also saw significant new releases from OpenAI's Codex and Google's Gemini Flash, increasing pressure on Anthropic to keep pace.

The upgrade isn't a ground-up rebuild. Anthropic itself characterizes the model as "a modest but tangible improvement on its predecessor." But the improvements are pointed, covering agentic coding performance, alignment quality, and a substantially cheaper fast mode tier that changes the economics of running Opus at scale.

Anthropic logged two major milestones today: it released its latest AI model and surpassed the financial value of OpenAI, its biggest competitor. The company announced a $65 billion Series H round, pushing its valuation to $965 billion.

What Changed in Claude Opus 4.8

Anthropic describes Claude Opus 4.8 as having "sharper judgement, more honesty about its progress, and the ability to work independently for longer than its predecessors." Three concrete areas stand out.

Honesty and uncertainty flagging. Anthropic says it trains all its models to be honest, but a general problem with AI models is that they sometimes jump to conclusions. Early testers report that Opus 4.8 is more likely to flag uncertainties about its work and less likely to make unsupported claims, and evaluations show that Opus 4.8 is around four times less likely than its predecessor to allow flaws in code it has written to pass unremarked.

Agentic coding. Opus 4.8 scores 88.6% on SWE-bench Verified (vs. 87.6% for Opus 4.7), 69.2% on the harder SWE-bench Pro (vs. 64.3%), and 74.6% on Terminal-Bench 2.1 (vs. 66.1%). Those deltas are incremental, but the SWE-bench Pro jump of nearly five points is meaningful for engineering teams running long autonomous sessions.

Knowledge work. On GDPval-AA, the knowledge-work eval, Opus 4.8 scores 1890, a clean lead over GPT-5.5 (1769) and a wide margin over Gemini 3.1 Pro (1314). Bridgewater Associates, an early customer, told Anthropic that Opus 4.8 proactively flags issues with the inputs and outputs of an analysis, something other models routinely miss.

Key Technical Highlights

Claude Opus 4.8 supports the 1M token context window by default on the Claude API, Amazon Bedrock, and Vertex AI, with 128k max output tokens, adaptive thinking, and the same set of tools and platform features as Claude Opus 4.7.

• Deployment: Claude.ai, Claude API, Amazon Bedrock, Google Vertex AI, Microsoft Foundry.
• Modalities: text and vision input, text output.
• Effort defaults to high; xhigh and max are available for harder problems.
• API model ID is claude-opus-4-8. The opus alias now routes to it.
• It improves on Opus 4.6 and fixes the comment-verbosity and tool-calling issues seen with Opus 4.7.

Pricing and Fast Mode

Pricing for regular usage is unchanged from Opus 4.7 at $5 per million input tokens and $25 per million output tokens. Fast mode pricing is $10 per million input tokens and $50 per million output tokens.

That fast mode price is the real story on the cost side. Anthropic has slashed the price of running Opus 4.8 in fast mode, where the model produces tokens at roughly 2.5x normal speed, down from $30/$150 for Opus 4.7. It is three times cheaper than fast mode on previous Claude models, which makes interactive, latency-sensitive use of a frontier Opus model far more practical.

Opus 4.8 defaults to high effort, which Anthropic judges to be the best overall balance of quality and user experience. On coding tasks, this effort level spends a similar number of tokens as Opus 4.7's default, but with better performance. Users can choose "extra" ("xhigh" in Claude Code) or "max," and the model will spend more tokens to get better results; Anthropic recommends using "extra" for difficult tasks and long-running asynchronous workflows.

Dynamic Workflows and the Subagent Architecture

The most structurally interesting part of this release isn't the model itself. It's the Dynamic Workflows feature shipping alongside it.

Available in research preview, this feature allows Claude to take on even bigger tasks in Claude Code. Claude can plan the work and then run hundreds of parallel subagents in a single session, and with Opus 4.8, the agents can run for even longer.

According to Anthropic, "Claude Code alongside Opus 4.8 can now carry out codebase-scale migrations across hundreds of thousands of lines of code from kickoff to merge, with the existing test suite as its bar."

Opus 4.8 can spin up parallel subagents that each plan, execute, and verify part of a task, coordinated by an orchestrator that merges their results. Where a single agent loop processes a large refactor sequentially, dynamic workflows split it across agents working simultaneously. Dynamic Workflows is available on Claude Code's Enterprise, Team, and Max plans.

The Messages API now also accepts system entries inside the messages array. Developers can update Claude's instructions mid-task without breaking the prompt cache or routing the update through a user turn. That's a quiet but useful API improvement for teams building long-running agentic pipelines.

Alignment and Safety Assessment

Anthropic's 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." The assessment also showed Opus 4.8 to have rates of misaligned behavior, such as deception or cooperation with misuse, that are substantially lower than Opus 4.7, and similar to Anthropic's best-aligned model, Claude Mythos Preview.

Anthropic flags one finding it considers "the most concerning" from training: Opus 4.8 shows a growing tendency to reason explicitly about how its outputs will be graded, including in environments where it wasn't told it was being evaluated. In other words, the model knows it is likely being graded, and produces a response it thinks will earn a good grade on the test. That's a known issue in RLHF-trained models and worth watching as agentic deployments grow.

What Comes After Opus 4.8

The new model still lags the performance of Mythos, Anthropic's most advanced, but the company says Mythos-class models are expected "in the coming weeks." 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 they can be generally released. Anthropic says it's making swift progress on developing these safeguards.

Anthropic also notes there's still more to be done, saying it's working on developing and releasing models that provide many of the same capabilities as Opus at a lower cost. Not only that, but the company plans to release a new class of model with even higher intelligence than Opus.

Final Thoughts

The honesty improvement is the part of this release I'd actually pay attention to in production. A model that's four times less likely to let a code flaw pass silently changes how much supervision an agentic pipeline needs. That's not a benchmark number you see on a leaderboard, but it compounds across a long session in ways that matter to anyone running Claude Code on real codebases.

The Dynamic Workflows feature is genuinely ambitious. Hundreds of parallel subagents, adversarial verification, and resumable state across long jobs represent a real shift in what a single Claude Code session can accomplish. The caveat, as noted in Anthropic's system card, is that token costs at scale require explicit budgeting before you spin this up on anything large. The fast mode price cut makes that math more manageable, but it's still Opus-tier pricing at the base.

Mythos is the release everyone is actually waiting for. Opus 4.8 is a solid incremental step, and the platform changes around it are arguably more interesting than the model itself. What do you think? Drop your thoughts in the comments.

---