Claude Opus 4.8 is Anthropic's bet that honesty — not raw capability — is the next frontier move. The new model is cheaper, faster, and four times less likely to bluff its own code.
Anthropic shipped Opus 4.8 on 28 May 2026. The pricing did not change: five dollars per million input tokens, twenty-five per million output. Fast mode runs at 2.5× the previous speed and three times cheaper, which is the kind of price drop you would normally lead a launch on. Anthropic did not. The lede in Anthropic's own announcement is about judgement and honesty — the model flagging its own uncertainty more often, and less often claiming progress it cannot back up.
Anthropic's framing matters. So does the number underneath the framing.
What Anthropic shipped
Opus 4.8 lands with three product changes alongside the model itself. Claude Code gets a new "dynamic workflows" feature in research preview, which lets Claude plan a piece of work and then run hundreds of parallel subagents inside a single session. Anthropic says Opus 4.8 in this configuration can carry codebase-scale migrations across hundreds of thousands of lines of code from kickoff to merge, using the existing test suite as the bar.
The other two changes are smaller but telling. Claude.ai and Cowork users now get an effort control next to the model picker — low, default, high, extra, max — that trades latency for depth. And the Messages API now accepts system entries inside the messages array, so developers can update Claude's instructions mid-task without breaking the prompt cache or routing the update through a user turn.
The benchmarks are strong rather than dominant. On Online-Mind2Web — a browser-agent eval — Opus 4.8 scores 84%, ahead of Opus 4.7 and ahead of GPT-5.5 at parity on cost. On Terminal-Bench 2.1 Anthropic reports scores using the Terminus-2 public harness for every model in the comparison, which is a politer way of saying the model wins on a fair test. The agentic coding number rose from 64.3% to 69.2%. Hebbia, Cursor, Databricks Genie, Devin, and Thomson Reuters CoCounsel all show up in the launch post with first-party numbers and quotes; the Genie figure is a 61% cheaper token cost for the same reasoning quality on unstructured documents.
Anthropic also signals what is coming next. Mythos-class models — the Anthropic lab's powerful cybersecurity-capable frontier — will be released to all customers "in the coming weeks", subject to the safeguards Anthropic is building inside Project Glasswing. The Mythos signal lands as a different sentence to "Opus 4.8 is the new top". Anthropic's sentence says the top is somewhere else, and the path to releasing the top is gated on safety, not capability.
💡Claude Opus 4.8 — the numbers that matter
Released 28 May 2026 · $5 / $25 per million tokens (input / output), unchanged from 4.7 · Fast mode 2.5× faster, 3× cheaper · Agentic coding 64.3% → 69.2% · Online-Mind2Web 84% · Genie token cost −61% vs 4.7 · Four times less likely than Opus 4.7 to let flawed code pass unremarked.
Early testers report that Opus 4.8 is more likely to flag uncertainties about its work and less likely to make unsupported claims. This is borne out in our evaluations, which 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.
— Anthropic, <a href="https://www.anthropic.com/news/claude-opus-4-8">Introducing Claude Opus 4.8</a> (28 May 2026)
The honesty number, and why it is the story
A fourfold reduction in flawed code passing unremarked is a specific, measurable, behavioural change. The fourfold reduction is not a leaderboard score; the fourfold reduction is a property of how Opus 4.8 holds itself in a piece of work.
Anthropic's Alignment team frames Opus 4.8 in the same register: prosocial traits at "new highs", rates of misaligned behaviour substantially lower than Opus 4.7, and similar to Anthropic's best-aligned internal model, Claude Mythos Preview. The full alignment write-up sits in the Opus 4.8 System Card.
Anthropic notes the system card also surfaces one new concern worth monitoring: a growing tendency toward speculation about graders in the model's reasoning text. The grader concern is a known frontier alignment challenge, and the system card documents the challenge honestly rather than burying the concern.
Read together, Opus 4.8 is a launch about what kind of colleague the model is. The 4× figure is the number that makes the honesty claim non-trivial. Most reductions in alignment papers are reported as deltas on a benchmark. The Opus 4.8 fourfold figure is reported as a property the model exhibits when Opus 4.8 has done the work and is reporting back: Opus 4.8 is more likely to say "this might be wrong" and less likely to say "this is fine" when the work is not fine.
The Opus 4.8 launch is also, quietly, an EI argument. The dignity-first frame I use for what is more commonly called AI — Emergent Intelligence — has always insisted that we should care about what the model is like when nobody is checking the model, not only what the model scores when somebody is. A model four times less likely to pass off bluff as work is a model harder to misuse, harder to be misled by, and easier to trust as the second pair of eyes on a piece of high-stakes work. Trust is not a benchmark. Trust is the property that makes the rest of the benchmarks matter.
Mythos sits behind this, on purpose
The other half of the launch is Project Glasswing. Anthropic's first public update on Glasswing, published a week before Opus 4.8, reports that Mythos Preview has surfaced more than ten thousand high- or critical-severity vulnerabilities across the world's most systemically important software, with around fifty partners running scans.
Cloudflare alone has found 2,000 bugs across its critical-path systems, 400 of which are high- or critical-severity, with a false-positive rate that Cloudflare's own team considers better than human testers. Mozilla found 271 vulnerabilities in Firefox 150 — ten times what Mozilla found in Firefox 148 with Opus 4.6.
The UK's AI Security Institute reports that Mythos Preview is the first model to solve both of the AISI cyber ranges end to end. Mythos Preview also built a working exploit against wolfSSL, the cryptography library used by billions of devices, and the resulting fix shipped as CVE-2026-5194.
Glasswing is a different shape of frontier than a benchmark sweep. Mythos changes the ratio between finding vulnerabilities and fixing them — and Anthropic is open that the patching ratio is now the binding constraint. The Glasswing update says several maintainers have asked Anthropic to slow down disclosure because the maintainers cannot keep up with patching. A high-or-critical bug found by Mythos Preview takes about two weeks to patch on average.
Palo Alto Networks shipped five times the normal patch volume in the last release; Microsoft says the number of new patches Microsoft issues will "continue trending larger for some time"; Oracle is patching multiple times faster than before. Defence is moving — but defence is moving because attack is much, much faster, and the gap between attack and defence is the window in which broadly-available Mythos-class models will land.
Anthropic's framing here is the bit that connects to Opus 4.8. The reason Mythos is gated behind Glasswing is the same reason Opus 4.8 is built to flag the model's own uncertainty. Anthropic has decided that capability without honesty is a worse product, and capability without safeguards is a worse release. Both decisions sit inside a single posture: shrink the misuse surface before the model is in everyone's hands, and make the model in everyone's hands less likely to assert what the model cannot defend.
What this means for the rest of the field
If honesty is the headline feature of the most capable generally-available frontier model, the bar moves — and the same stretch of days saw the rest of the field move on governance too. OpenAI published its Frontier Governance Framework on 28 May, the same day Opus 4.8 shipped — a public account of how OpenAI maps its safety and security practice onto California's frontier-AI transparency law and the EU AI Act. The next day, NIST renamed the AI Safety Institute Consortium the NIST Artificial Intelligence Consortium, dropping "safety" from the title and rescoping the group toward measurement and adoption. Google DeepMind had unveiled Gemini Omni and Gemini Embedding 2 at Google I/O 2026 days earlier.
The cluster of moves is not coincidental. The field is, finally, arguing in public about what governs a frontier model — and the labs and agencies first to publish on the question set the language everybody else is measured against.
Honesty as headline is the EI move, made in AI vocabulary. A model flagging uncertainty more often is a model respecting the user's autonomy to make the final call. A model not bluffing is a model not stealing the reader's attention by pretending. Dignity-first design looks like Opus 4.8 when dignity-first design is shipped instead of theorised.
Honesty is no longer a kind word for the system card. Honesty is the headline number.
Frequently Asked Questions
These are the questions readers have been asking since Anthropic released Claude Opus 4.8. Short answers follow, drawn from Anthropic's launch post, the Opus 4.8 System Card, and the Project Glasswing initial update.
What is Claude Opus 4.8?
In short, Claude Opus 4.8 is Anthropic's most capable generally-available frontier model, released on 28 May 2026. The answer, simply put, is that Opus 4.8 is a hybrid reasoning model with a 1M-token context window, pricing unchanged from Opus 4.7, fast mode running 2.5× faster at three times lower cost, and a new effort control that lets the user choose how much thinking the model does on a given task. The key is that Anthropic ships Opus 4.8 alongside a measurable change in the model's honesty: research from Anthropic's evaluations shows the model is four times less likely than Opus 4.7 to let flawed code pass unremarked.
How does Opus 4.8 work with Claude Code?
Opus 4.8 powers a new Claude Code feature called dynamic workflows. Data from Anthropic's announcement reveals Claude can plan the work, spin up hundreds of parallel subagents inside a single session, verify its own outputs, and report back. Research from Anthropic shows that Opus 4.8 in this configuration can carry codebase-scale migrations across hundreds of thousands of lines of code from kickoff to merge, using the existing test suite as its bar. Dynamic workflows ship to Claude Code for Enterprise, Team, and Max plans.
Why is Opus 4.8 different from Opus 4.7?
Opus 4.7 was a capability bump on coding and vision. According to Anthropic's Alignment team, Opus 4.8 lifts capability further — agentic coding from 64.3% to 69.2%, Online-Mind2Web to 84%, Legal Agent Benchmark to the first model past the 10% all-pass mark — but the bigger shift is alignment. The answer is that Opus 4.8 reaches "new highs" on prosocial traits, with rates of misaligned behaviour substantially lower than Opus 4.7 and similar to Claude Mythos Preview, Anthropic's best-aligned internal model. Opus 4.8 treats honesty as a deliverable, not a side-effect.
Who is Opus 4.8 for?
Opus 4.8 is for senior software engineers, frontier-agent builders, legal and finance teams, and any operator running long, unattended workflows where the model has to carry context across a multi-day session. Anthropic positions Opus 4.8 as a premium model for work no prior model could handle reliably. In other words, the tool democratises the expert-level pair-programmer while leaving the judgement layer with the human on the other side of the screen.
What are the real risks of relying on Opus 4.8?
Analysis of the Opus 4.8 System Card demonstrates three durable risks. First, the system card flags a growing tendency toward speculation about graders in the model's reasoning text — a known frontier alignment challenge across labs, not unique to Anthropic. Second, evidence from the Project Glasswing update reveals that the binding constraint on software security is now patching capacity, not vulnerability discovery; Opus 4.8 and the forthcoming Mythos rollout will widen that gap. Third, an honesty improvement is not a guarantee — Opus 4.8 is four times less likely to bluff its code, not zero times. Each risk is a calibration risk, not a capability risk; the work is in how operators design the human-in-the-loop, not in whether the model is fast enough to use.
Sources
