Anthropic shipped Claude Opus 4.8 roughly 41 days after Opus 4.7, which is an unusually tight turnaround for a frontier vendor and tells you most of what you need to know about the release. This is a refinement, not a reset. Pricing held steady at $5 per million input tokens and $25 per million output tokens, the 1M-token context window stuck around on the main platforms, and the headline gains went to coding, browser-agent reliability, and tool-calling behaviour. If you were already running Opus 4.7 in production, the move to 4.8 is a low-friction upgrade. If you were not, the more interesting question is where Opus 4.8 sits inside Anthropic’s broader model hierarchy, because it is not the most capable thing Anthropic has built.
Quick answer: what changed in Opus 4.8
As of May 28, 2026, Claude Opus 4.8 is Anthropic’s generally available flagship for coding, agent work, and professional analysis. Compared with 4.7 it ships stronger coding and reasoning, better browser and computer-use performance (including 84% on Online-Mind2Web), fixes for verbose comment generation, and cleaner tool-calling. Pricing is unchanged. Context windows are unchanged. Anthropic frames the model as more honest and better at judgement, which in practice means it speculates less and stays inside instructions more reliably.
“More honest behavior and better judgement” reads as marketing copy until you watch it acknowledge missing information instead of papering over a gap. That is the actual upgrade.
What Anthropic actually shipped
The model card and announcement focus on three things: code quality, agentic reliability, and trust behaviour. Anthropic’s own description positions Opus 4.8 as an improvement on long-running work, not as a benchmark statement. That matters because the failure modes in real agent deployments are rarely about raw IQ. They are about tools called with the wrong arguments, output drowning in commentary, loops that will not terminate, and models that confidently invent context they should have flagged as missing.
Opus 4.8 chips away at all of those.
What could a custom AI agent take off your plate?
We build production-grade AI systems that quietly handle the busywork, so your team can focus on the work that actually matters.
The browser-agent number deserves a closer look. An 84% on Online-Mind2Web is a state-of-the-art result, but the more useful framing is that browser tasks demand precise action selection, state tracking, and recovery under ambiguity. Those are exactly the skills you need for any long-horizon agent, not just web navigation. If 4.8 has gotten visibly better at clicking the right thing on a complex page, it has probably also gotten better at choosing the right database query, the right shell command, and the right next step in a multi-tool pipeline.
The third change, around honesty, is harder to measure but easy to notice in use. Earlier Opus generations would sometimes confidently fabricate a config flag or a function signature. The reporting from Trending Topics’ launch coverage frames 4.8 as more willing to say “I do not have this information” or to ask for clarification before acting. For anyone running it in an agent loop, that is worth more than a few benchmark points.
Pricing, context windows, and the thing that quietly costs you more
Here is where the story gets a bit awkward. The list price has not moved in three releases.
| Model | Input per 1M tokens | Output per 1M tokens | Context window (API, Bedrock, Vertex) |
|---|---|---|---|
| Opus 4.6 | $5 | $25 | 1M |
| Opus 4.7 | $5 | $25 | 1M |
| Opus 4.8 | $5 | $25 | 1M |
That looks like price stability. On effective cost, it is not quite the same picture. The tokenizer changes that arrived with the 4.7 generation map the same text to roughly 1.0 to 1.35 times more tokens than earlier Claude generations, with multilingual content and heavily structured data sitting toward the higher end of that range. So if your workload is JSON-heavy, code-heavy, or non-English, you are probably paying meaningfully more per task than the unchanged price card suggests. This is the kind of detail you only notice after the first invoice.
Two more practical notes. On Microsoft Foundry, Opus 4.8’s context window is capped at 200K tokens rather than the 1M you get on Anthropic’s API, Bedrock, or Vertex. If you have been architecting for the full million and you are routing through Foundry, that is a constraint worth confirming before you deploy. And the task budgets feature (in beta) lets you cap total tokens across an entire agent loop, including thinking, tool calls, and outputs. The model can see its own remaining budget and adjust. For anyone who has watched a runaway agent burn through $40 of inference before someone noticed, this is the single most useful piece of operational scaffolding in the Claude stack.
Is Opus 4.8 actually Anthropic’s most powerful model?
No. And this is the part of the launch that most coverage has glossed over.
The supplied benchmarks point clearly to Mythos Preview, an Anthropic model restricted to Project Glasswing partners and focused on defensive cybersecurity, as the more capable system overall. The comparison data from NxCode reports Mythos at 93.9% on SWE-bench Verified and 83.1% on cybersecurity benchmarks, against Opus 4.7’s 80.8% and 66.6%. The detailed line-by-line synthesis suggests the Mythos advantage is largest on coding, long-horizon agent tasks, and cyber, and narrows toward statistical noise on perception and graduate-level reasoning.
The implication is straightforward. Anthropic is running a segmentation strategy:
- Opus 4.8 is the broadly available, high-trust model for coding, agents, and professional work.
- Mythos is the more capable system, kept inside a restricted partner programme aimed at defensive security.
- Opus 4.7 remains around as the previous generation for teams that have not yet migrated.
That is a deliberate split between “maximum capability” and “widely deployable capability”. For most organisations the second matters more, and Opus 4.8 is built squarely for that audience. But if you are evaluating Claude against frontier offerings from other vendors, you are not seeing Anthropic’s strongest hand.
Curious what AI could do for your business?
No jargon and no hard sell. Just a friendly look at where AI fits, and where it doesn't.
What the Claude Code leak revealed about how Opus 4.8 really works
In a separate incident, Anthropic accidentally exposed roughly 512,000 lines of Claude Code TypeScript source, reportedly because of a missing .npmignore. The New Stack’s writeup of the leak is the cleanest account of what was inside.
This is not strictly about Opus 4.8. But it is about the production environment Opus 4.8 lives in, and that environment matters more than people realise.
A few things stood out:
- Multi-agent swarms with deliberately restricted toolsets per sub-agent.
- Persistent memory through plain inspectable files (
CLAUDE.md), not a vector database. - Tool registries with JSON schema validation, permissioning, and risk classification.
- Background daemons and session managers for long-running work.
The takeaway: the model is one part of a system designed to keep it bounded, auditable, and recoverable. The reason Opus 4.8’s tool-calling fixes and verbosity reductions matter so much is that they slot into a scaffolding stack that already does compaction, budget enforcement, and sub-agent dispatch. A better model inside this system is worth more than a better model running raw. If you have ever tried to build agent infrastructure from scratch, you know the gap between “the model is smart” and “the system stays up overnight” is enormous, and it lives in exactly these primitives.
What to watch for if you are deploying it
A few practical points, drawn from the supplied research and from the patterns that show up across the Opus 4.7 and 4.8 generations.
- Be more explicit in prompts than you think you need to be. Opus has gotten noticeably more literal across the last two releases. It will do exactly what you ask, which means missing context tends to produce stuck behaviour rather than helpful inference. Define scope, constraints, stop conditions, and tool contracts.
- Budget for tokenizer drift. Same price, possibly more tokens per task. Multilingual and structured workloads are where this bites hardest.
- Treat Opus 4.8 as a reasoning core, not a renderer. It cannot natively cut video, build files, or perform external actions. Pair it with execution tools and keep the boundaries clean.
- Use task budgets. Especially if you are running anything that looks like an agent loop. The cost-of-a-mistake floor drops dramatically.
- Do not assume you are seeing Anthropic’s best. If your evaluation question is “how does Claude compare to model X on cyber or long-horizon code”, Opus 4.8 is not the right comparison point. Mythos is, and you probably cannot access it.
How to decide whether to upgrade
If you are on Opus 4.7 and your workload involves multi-tool coding agents, browser automation, or document-heavy analysis, the move to 4.8 is worth doing this quarter. The pricing is unchanged, the context window is unchanged on the platforms that matter, and the verbosity and tool-calling fixes alone repay the migration effort for most teams. If you are on a competing model and shopping, evaluate 4.8 on your actual workflow rather than on headline benchmarks, and pay attention to the operational features (task budgets, compaction, large context) as much as raw model behaviour. The honest reading of this launch is that Anthropic is selling reliability, not novelty, and reliability is what production work actually needs.
