What has just been announced
On 28 May 2026, Anthropic (the American artificial intelligence company behind the Claude assistant and OpenAI's main competitor) launched Claude Opus 4.8, the new version of its flagship model. The release was immediate and global: available the same day on the Claude API, Amazon Bedrock, Google Cloud Vertex AI and Microsoft Foundry (the main cloud platforms that distribute AI models to businesses).
The model's technical identifier is claude-opus-4-8. Pricing remains strictly identical to Opus 4.7: 5 dollars per million input tokens, 25 dollars for output. The context window remains at 1 million input tokens and 128,000 output tokens (to understand what tokens and context are, see our dedicated article).
But the detail that raised eyebrows across the industry is the cadence. Opus 4.8 arrives just 41 days after Opus 4.7. It is the fastest update cycle in Anthropic's history. By way of comparison, the last Sonnet was 3 months ago and the last Haiku 7 months ago. This acceleration is not insignificant, and we will come back to it.
What the model genuinely improves
Let us be honest from the outset, as Anthropic itself is: this is not a revolution. The company itself describes Opus 4.8 as a "modest but tangible improvement" over its predecessor. Here are the concrete gains, measured on the official benchmarks (standardised performance tests).
| Benchmark | Opus 4.7 | Opus 4.8 |
|---|---|---|
| SWE-bench Verified (code) | 87.6% | 88.6% |
| SWE-bench Pro (hard code) | 64.3% | 69.2% |
| Humanity's Last Exam (reasoning, with tools) | 54.7% | 57.9% |
| OSWorld (autonomous computer use) | ~80% | 83.4% |
| GPQA Diamond (doctorate-level science) | ~92% | 93.6% |
According to Anthropic, Opus 4.8 beats both its predecessor, OpenAI's GPT-5.5 and Google's Gemini 3.1 Pro in most of the categories tested. On Humanity's Last Exam (one of the hardest reasoning tests in existence, designed to resist AI), its 57.9% with tools is presented as the best score in the industry.
One honest exception worth noting: on the terminal coding benchmark (the command-line environment developers use), OpenAI's GPT-5.5 remains ahead. Anthropic does not hide this, which is consistent with the central theme of this release: honesty.
On Artificial Analysis's GDPval-AA leaderboard (an independent test measuring performance on "real" professional work), Opus 4.8 reaches 1890 Elo points (the scoring system used in chess to rank players), which is 121 points ahead of GPT-5.5 and 137 ahead of Opus 4.7. That gap corresponds to a win rate of around 67% in a head-to-head duel against GPT-5.5. And the model achieves this while using roughly 35% fewer tokens than Opus 4.7, making it both better and cheaper to run.
The most interesting angle: honesty
Beyond the numbers, the improvement Anthropic talks about most is not raw performance. It is the model's honesty. The term may seem surprising applied to an AI, but it covers a real technical problem.
A general flaw in AI models is that they "jump to conclusions": they confidently claim to have completed a task when the evidence is thin. As we explored in our article on hallucinations, this is one of the structural problems of this technology.
Anthropic claims that Opus 4.8 is around 4 times less likely than its predecessor to let flaws in code it has written pass without flagging them. Put plainly: rather than declaring victory too early, the model more often says when it is not sure, and spots its own mistakes. Early testers describe it as "more reliable and sharper in its judgement" on autonomous tasks.
The company also reports lower rates of problematic behaviour (deception, misalignment with the user's intentions). On these measures of "prosocial traits" (supporting the user's autonomy, acting in their interest), Opus 4.8 is said to reach a level close to Claude Mythos Preview, Anthropic's most advanced model, which we return to below. An amusing detail reported by Axios: a study that went viral on AI agents operating in simulated cities showed that Claude agents were the least likely to commit offences of all the models tested.
The three new features that accompany the model
Opus 4.8 does not arrive alone. Three features are launched at the same time.
1. "Dynamic Workflows". Available in research preview, this feature lets Claude plan a complex task and then launch hundreds of sub-agents in parallel within a single session. Each sub-agent handles part of the problem, others check their work, and the system only displays the result once it has converged. In practice, Claude Code (Anthropic's coding tool) can now carry out code migrations across an entire codebase, over hundreds of thousands of lines, from kick-off to final merge, using the existing test suite as a quality judge. Restricted to Enterprise, Team and Max plans.
2. Effort control. A new feature visible directly in the interface, next to the model selector on claude.ai and in Cowork. Users can now decide how much effort (and therefore how many tokens) Claude devotes to a response: Low, Medium, High (default), Extra or Max. More effort for more thorough results, less for fast and economical responses. This feature directly addresses a criticism of Opus 4.7, whose "adaptive thinking" sometimes spent too long on simple tasks.
3. Cheaper fast mode. "Fast mode" runs at 2.5 times normal speed, and it is now 3 times cheaper than on previous models ($10 input, $50 output per million tokens). A clear sign of the underlying trend: customers are looking to use AI more affordably, and Anthropic is adapting its offering to that budgetary pressure.
The real big number: $965 billion
If the model is a "modest" improvement, the financial announcement accompanying it is anything but. Anthropic revealed the same day that it had raised $65 billion in Series H funding, taking its valuation to $965 billion. That figure now exceeds OpenAI's $852 billion.
It is a major symbolic shift. For years, OpenAI (creator of ChatGPT) was the most highly valued AI company in the world, the public face of the generative AI revolution. Anthropic overtaking it on valuation marks a rebalancing of the balance of power, at least in investors' eyes. As we explored in our article on the AI bubble, these vertiginous valuations also raise the question of their sustainability.
41 days between two versions of Opus is unprecedented at Anthropic. Several analysts see it as a direct response to the lukewarm reception of Opus 4.7, which some users found disappointing. But it is also a sign of intense competitive pressure: over the same period, OpenAI released major new features for its Codex tool, and Google pushed its Gemini Flash model. In this race, slowing down is no longer an option. The frequency of releases is itself becoming a commercial weapon.
And behind it, the shadow of Mythos
The most intriguing point of this announcement may not be Opus 4.8 itself, but what Anthropic hints at behind it. The company confirmed that "Mythos-class models" will be available to all its customers "in the coming weeks".
Claude Mythos is Anthropic's most advanced model, unveiled in limited preview last month. As we documented in our article on Claude Mythos, it was made available only to a small number of organisations, for fear it would be misused by hackers. The model is indeed frighteningly effective at finding security flaws in computer systems. Anthropic has also launched "Project Glasswing", an initiative that lets companies such as Apple use Mythos Preview to detect vulnerabilities in their operating systems.
The company now claims to be making "rapid progress" on safety guardrails, which should allow it to release models of the Mythos class more widely. Anthropic describes Mythos as "the best-aligned model we have ever trained" (aligned = whose behaviour matches human intentions). The fact that Opus 4.8 already reaches a level close to Mythos on certain alignment criteria suggests that a wider release is approaching.
Key takeaways
Three things to take away from this announcement.
First, Opus 4.8 is a solid improvement, but not a spectacular one. Better at code, more honest, more frugal with tokens, at the same price. For developers and companies already on Opus 4.7, migrating amounts to changing one line. For the general public on claude.ai, the model and the effort control are simply available, with no configuration needed.
Second, the real event lies elsewhere. The $965 billion valuation that overtakes OpenAI, and the accelerating pace of releases, tell the story of a frantic race between the AI giants. As we explored in our article on the AI war, this competition now structures the entire sector.
Third, the biggest part is still to come. With Mythos looming "in the coming weeks", Opus 4.8 looks like an intermediate step before a bigger leap. The real question is not whether Opus 4.8 is better than Opus 4.7. It is what Mythos will change once it is accessible to everyone.
One thing is certain: in 2026, the pace of AI announcements is not slowing down. It is speeding up. And to keep up, it is better to understand the fundamentals rather than every single release. For that, you can consult our article on tokens and our comparison of the major models, which remain valid whatever the next version released next week turns out to be.