Claude Opus 4.8 Guide: The Opus 4 Model That Outlived Its Successors

Opus 4.8 arrived on 28 May 2026, six weeks after Opus 4.7, and was described at launch as a modest but clear improvement. Two newer Opus models have shipped since. Yet 4.8 is not the forgotten middle child it looks like. Anthropic now routes most cybersecurity requests flagged on Opus 5 and Opus 5.5 back to it, which makes it part of the serving path for its own successors.

That is the useful frame for this model. It is the last of the Opus 4 line, the one with the fewest moving parts, and the one Anthropic itself treats as the safe landing spot.

A drop-in upgrade with no API changes

Opus 4.7 was the release that broke integrations: it removed thinking budgets and sampling parameters and changed the tokenizer. Opus 4.8 kept that request surface exactly. Anthropic's migration notes describe a 4.7 to 4.8 move as a model-ID swap plus prompt re-tuning, with nothing else required to avoid an error.

Pricing also stayed put at $5 per million input tokens and $25 per million output, with a 1M-token context and 128K maximum output at standard rates. The additions were operational:

• Mid-session system messages. You can place role: "system" entries inside the messages array to change instructions partway through a task without invalidating the prompt cache.
• Cheaper fast mode. Anthropic priced fast mode at $10 / $50 per million tokens and said it runs at about 2.5 times normal speed, at a third of what fast mode cost on earlier models.
• Effort control in the consumer apps, defaulting to high.

If you came from Opus 4.6 or earlier, the breaking changes are the ones the Opus 4.7 guide lists. They apply unchanged.

Honesty was the headline, not the scores

The most interesting claim in the launch was behavioural. Anthropic says Opus 4.8 is about four times less likely than Opus 4.7 to let flaws in its own code pass without comment, and early testers reported it flagging uncertainty more often and making fewer unsupported claims.

For agentic coding this matters more than a benchmark point. The expensive failure in an unattended loop is not a wrong answer; it is a wrong answer reported as done. A model that says "this test is flaky, I could not confirm the fix" saves a review cycle that a confident model costs you.

There is a caveat in Anthropic's own guidance. Opus 4.8 follows review instructions literally. If a review prompt says "only report high-severity issues", measured recall can drop even though underlying bug-finding improved. Ask it to report everything and filter downstream. Best model for code review covers why precision and recall pull in different directions here.

The benchmark picture

Anthropic's comparison table was an image, so most numbers circulated through secondary coverage of it. As reported there:

• SWE-bench Pro — 69.2%, up from 64.3% for Opus 4.7
• Terminal-Bench 2.1 — 74.6%, behind GPT-5.5 on that evaluation
• OSWorld-Verified — 83.4%
• Humanity's Last Exam — 49.8% without tools, 57.9% with tools

On GDPval-AA v2, the professional-work benchmark, Opus 4.8 sits at 1,600 against 1,815 for Claude Fable 5 Max. All of these are vendor-chosen evaluations on vendor scaffolds; treat them as a direction of travel rather than a forecast of your results.

Why 4.8 outlives its successors

When Opus 5 launched in July, Anthropic said requests flagged by its cyber classifiers fall back to Opus 4.8 by default. Opus 5.5 kept the arrangement: most cybersecurity tasks are rerouted to 4.8 under its safeguards. Anthropic's documentation describes 4.8 as the recommended target for cyber-category refusals, so a fallback genuinely recovers the request rather than relabelling the failure.

Two practical consequences follow:

1. If you run security tooling, some of your Opus 5 or 5.5 traffic may already be served by 4.8. Calling 4.8 directly for that work removes a classifier round-trip and makes the behaviour predictable.
2. If you build your own fallback chain, 4.8 is a sensible second model behind a newer Opus. It shares the request surface, the context size and the price tier. Fallback model selection covers what makes a good second model.

Anthropic lists its retirement as not sooner than 28 May 2027.

Prompt behaviour that shifted

None of these break code, but prompts tuned for 4.7 may land differently. Anthropic's notes describe 4.8 as:

• More talkative between tool calls and at the end of a task. Remove any "summarise every few steps" scaffolding, and add a silence default if a coding agent narrates too much.
• More deliberate. It asks about small decisions it previously just made. Grant explicit autonomy on naming and defaults while keeping caution on destructive actions.
• Conservative with tools. It reaches for search, subagents and memory files less readily unless the prompt says when each applies.
• Warmer in prose, roughly the opposite direction from 4.7's clipped style. Style prompts added to counter 4.7 may now overcorrect.

When to choose Opus 4.8 today

Opus 5.5 is cheaper per token and stronger on Anthropic's published numbers, so 4.8 is no longer the default pick for new work. It still earns a place in three situations:

• Security research and tooling, where newer Opus models route the work to it anyway.
• Integrations that rely on behaviour restricted or removed later — disabling thinking at any effort level, forced tool_choice, or Priority Tier, which Anthropic does not offer on Opus 5 or 5.5.
• A fallback slot behind a newer Opus in the same tier.

The catalogue also lists a :batch variant: the same model on the asynchronous Batch API at half price, $2.50 / $12.50 per million tokens. For anything else, start with Opus 5.5 and keep 4.8 as the model you can fall back to.