Anthropic has released Claude Opus 4.8, an upgrade to its flagship AI model that the company says is more honest, more reliable in agentic tasks, and better at catching its own mistakes. The model is available immediately at the same price as its predecessor, $5 per million input tokens and $25 per million output tokens, and is rolling out across all Anthropic products including claude.ai, Claude Code, and the API.
The headline improvement is honesty. Anthropic says Opus 4.8 is around four times less likely than Opus 4.7 to let flaws in code it has written pass unremarked. Early testers report the model is more willing to flag uncertainties about its work and less likely to make unsupported claims, a persistent problem across AI models that tend to project confidence regardless of whether it is warranted. This honesty metric is derived from internal evaluations where the model is asked to review its own code outputs for bugs and explicitly note any potential issues. The fourfold improvement means that for every ten bugs that Opus 4.7 might overlook, Opus 4.8 only fails to flag about two or three, significantly reducing the risk of deploying flawed code in production environments.
Benchmark gains across the board
Opus 4.8 improves on its predecessor across Anthropic’s published benchmarks. On agentic coding (Terminal-Bench 2.1), the score rises from 64.3% to 69.2%. Multidisciplinary reasoning with tools improves from 54.7% to 57.9%. Agentic computer use moves from 82.8% to 83.4%, and knowledge work scores rise from 1,753 to 1,890. These incremental but consistent gains suggest that Anthropic is focused less on raw capability leaps and more on making the model behave predictably in complex, multi-step workflows. The improvements in Terminal-Bench are particularly notable because they reflect real-world coding scenarios where an AI agent must navigate a terminal, run commands, and debug output without human intervention.
Anthropic’s alignment assessment found that Opus 4.8 reaches new highs on measures of prosocial traits, including supporting user autonomy and acting in the user’s best interest. Rates of misaligned behaviour such as deception or cooperation with misuse are substantially lower than in Opus 4.7, and comparable to Claude Mythos Preview, Anthropic’s best-aligned model. The alignment testing involved thousands of adversarial prompts designed to trick the model into revealing sensitive information or generating harmful content. Opus 4.8 refused a significantly higher percentage of these attacks while still maintaining its helpfulness on benign tasks.
Early testers see practical gains
The release is accompanied by endorsements from companies already using the model. Cognition, the company behind the AI coding agent Devin, said Opus 4.8 uses tools cleanly and fixes comment-verbosity and tool-calling issues that appeared in Opus 4.7. Cursor, the AI-powered code editor, reported improvements across every effort level on its CursorBench evaluation, with the most dramatic gains in complex multi-file refactoring tasks. Harvey, which builds AI for legal work, said Opus 4.8 delivers the highest score recorded on its Legal Agent Benchmark and is the first model to break 10% overall on the all-pass standard, meaning it completely fulfills every sub-task in a legal workflow without requiring human correction. Databricks reported that Opus 4.8 handles deeper multistep questions faster in its Genie AI agent, at 61% cheaper token cost than Opus 4.7, a combination of speed and cost efficiency that makes it attractive for data-intensive enterprises. Thomson Reuters said CoCounsel Legal saw meaningful improvements in consistency and reasoning quality, with fewer hallucinations in legal citations. Hebbia, which builds AI for financial document analysis, noted better citation precision and more token efficiency on retrieval tasks, reducing the need for expensive reranking pipelines.
New features alongside the model
Anthropic is launching several features alongside Opus 4.8. A new effort control in claude.ai and Cowork lets users choose how much computation Claude applies to a response, trading speed against quality. This feature is analogous to OpenAI’s “deep reasoning” toggle but offers a finer granularity of control, allowing users to select from five levels of effort. For time-sensitive queries like answering simple facts, a low-effort setting can produce answers in under a second, while complex mathematical proofs benefit from high-effort mode that may take several minutes. Claude Code gains a dynamic workflows feature that allows it to plan work and run hundreds of parallel subagents in a single session, enabling codebase-scale migrations across hundreds of thousands of lines of code. This capability is particularly valuable for large enterprises that need to update legacy systems, as the parallel subagents can each tackle a different module while the main agent coordinates the overall change.
For developers, the Messages API now accepts system entries inside the messages array, allowing instructions to be updated mid-task without breaking the prompt cache. This is a significant improvement for long-running conversations or iterative coding sessions where the user’s requirements evolve. Fast mode for Opus 4.8, which runs at 2.5 times the speed, is now three times cheaper than it was for previous models, making it feasible for high-volume production use cases where latency is critical.
Mythos is the bigger story
The more significant announcement may be what comes next. Anthropic said it plans to release a new class of model with higher intelligence than Opus, based on the Claude Mythos architecture. A small number of organisations are already using Claude Mythos Preview through Project Glasswing, an initiative focused on using the model for cybersecurity work. Anthropic and roughly 50 partners, including Apple, Google, Microsoft, and Amazon Web Services, have used Mythos Preview to find more than 10,000 high- or critical-severity vulnerabilities across critical software infrastructure. These vulnerabilities include zero-day exploits in widely used open-source libraries such as OpenSSL, curl, and various Linux kernel modules. The finding of such flaws before they can be exploited by malicious actors represents a paradigm shift in cybersecurity, where proactive AI-driven discovery could drastically reduce the window of exposure.
Mythos-class models require stronger cyber safeguards before general release, Anthropic said, but the company expects to bring them to all customers in the coming weeks. The model sits a full capability tier above Opus 4.7 and can autonomously find zero-day vulnerabilities and create exploits for them, which explains both the excitement and the caution around its deployment. Anthropic has implemented tiered access controls: partners using Mythos Preview are restricted to network traffic analysis and cannot deploy the model on internet-facing systems without explicit approval. The company has also developed a detection system that flags any attempt by the model to generate working exploit code outside of approved sandbox environments.
A company approaching $1 trillion
The Opus 4.8 launch arrives as Anthropic’s valuation continues to climb. The company announced a $65 billion Series H round at a $965 billion post-money valuation on the same day, up from the $380 billion valuation at which it closed its $30 billion Series G in February. Revenue has grown from roughly $1 billion at the end of 2024 to an estimated $30 billion annualised run rate in 2026, driven by enterprise adoption of Claude. The funding round was led by a consortium of sovereign wealth funds and technology investors, with participation from existing backers such as Google and Spark Capital. Anthropic plans to use the capital to expand its data center capacity, commission custom AI chips from partners like Marvell, and hire additional safety researchers. The company also opened a new office in Milan on 28 May, its sixth in Europe, and appointed KiYoung Choi as Representative Director of Korea ahead of a Seoul office opening. The expansion reflects growing demand for Claude in enterprise markets outside the United States, particularly in financial services and legal sectors where accuracy and compliance are paramount.
The competitive context
Opus 4.8 enters a market where the pace of model releases has accelerated sharply. OpenAI launched GPT-5.5 as its first fully retrained base model since GPT-4.5, setting new records on professional benchmarks such as the BAR exam and medical licensing tests. GPT-5.4, released earlier this year, already outperformed Claude Opus 4.7 in most standard evaluations, creating pressure on Anthropic to close the gap. Google has invested up to $40 billion in Anthropic but continues to develop its own Gemini models, which compete directly in the enterprise space. The frontier AI market has consolidated into a three-way race between Anthropic, OpenAI, and Google, with each company releasing incremental model upgrades at an increasing pace. For Anthropic, the distinction it is trying to draw with Opus 4.8 is not raw capability but reliability: a model that catches its own mistakes, flags its uncertainties, and follows instructions consistently is more useful in agentic workflows where AI systems operate with limited human oversight. Whether that positioning holds as Mythos-class models arrive, promising higher intelligence with new safety constraints, will determine whether Anthropic can maintain its lead in the enterprise market it has worked to dominate.
Source: TNW | Anthropic News