Claude Opus 4.8 (2026): Benchmarks, Price & Verdict

Claude Opus 4.8 is the most capable model in Anthropic's Opus 4 series, released 28 May 2026 for the hardest reasoning and agentic coding. It tops SWE-bench Pro at 69.2%, costs $5 per million input and $25 per million output tokens, and holds a 1-million-token context window. Claude Opus 5 has since replaced it at the same price, but 4.8 remains callable as a legacy model.

Benchmarks, pricing, and lineup position verified 9 August 2026 against Anthropic's model overview.

We ran this model across production coding work on this site for weeks. This guide covers what it is, where it leads, what it costs, and — now that a newer Opus exists — whether there's still a reason to run it. If you want the whole family mapped first, our Claude models explained guide lays out every tier.

What is Claude Opus 4.8?

The top of Anthropic's Opus 4 series, released 28 May 2026 and tuned for software engineering and high-stakes enterprise work rather than casual chat. It has since been superseded by Opus 5 at identical pricing.

It builds directly on Opus 4.7 with better judgment, cleaner tool use, and noticeably stronger agentic coding. Anthropic introduced it as an engineering-focused release.

You reach it as the model ID claude-opus-4-8 on the Claude API, or inside Claude for Pro, Max, Team, and Enterprise plans. It's also available on Amazon Bedrock (anthropic.claude-opus-4-8), Google Cloud, Microsoft Foundry, and GitHub Copilot, so most enterprise stacks can call it without leaving their existing cloud. Two things define the release: a 1-million-token context window by default and an adaptive thinking mode that lets the model reason before it answers.

Where it sits now matters as much as what it does. Above it are Claude Opus 5 — the current Opus, at the same $5/$25 — and the premium Claude Fable 5 at roughly double the price. For new work, Opus 5 is the default; 4.8 earns its place when you're pinned to it, have a deployment already validated against it, or need it as the recommended fallback when an Opus 5 safety classifier declines a request.

Benchmarks: where it leads

SWE-bench Pro at 69.2% — nearly five points clear of Opus 4.7 — with the largest gains on tasks demanding sustained reasoning across long context.

On paper the jump from 4.7 looks incremental; on the hardest tests it isn't. On Anthropic's Super-Agent benchmark it was the only model to complete every case end-to-end.

The pattern is clear: the biggest gains land on tasks that demand sustained reasoning — long-context tracking nearly doubled, and competition math jumped almost thirty points. For everyday questions you won't feel the difference. For multi-hour agent runs and dense codebases, you will.

One comparison worth knowing: Claude Sonnet 5, at $3/$15, actually beats this model on agentic terminal work, scoring 80.4 on Terminal-Bench 2.1 against 74.6. The Opus advantage is real but narrower than the tier gap suggests — see Claude Opus 4.8 vs Sonnet 5 for the full split.

Pricing and Fast Mode

$5 per million input tokens and $25 per million output — identical to Opus 4.5, 4.7, and the newer Opus 5. Four consecutive releases at the same rate.

Two levers cut that further: prompt caching saves up to 90% on repeated context, and the Batch API halves the bill for non-urgent jobs. Our Claude API pricing breakdown has the full math.

Fast Mode runs the model at roughly 2.5 times normal output speed for $10 input and $50 output per million tokens. It requires the beta messages endpoint, the fast-mode-2026-02-01 beta flag, and speed: "fast" as a top-level request parameter. It's a research preview on the Claude API only — not available on Bedrock, Google Cloud, or Microsoft Foundry. Note that Opus 4.7's Fast Mode has been removed entirely, so 4.8 and Opus 5 are the only models that support it.

Effort levels and adaptive thinking

Five effort levels — low, medium, high, xhigh, and max — trading latency for depth, with high as the default everywhere.

Set it inside output_config, not at the top level: output_config: {effort: "xhigh"}.

Thinking is separate and off by default on this model — a quirk that catches people, because Opus 5 reverses it and thinks unless you disable it. To make 4.8 reason before answering, set thinking: {type: "adaptive"} explicitly. One further gotcha: temperature, top_p, and top_k are rejected and return a 400, so you steer behaviour through prompting instead. Our extended thinking guide covers when reasoning tokens are worth paying for, and the context window explainer covers how far 1M tokens actually stretches.

When to use it instead of Sonnet

When a job runs unattended for hours, spans a dense legacy codebase, or fails repeatedly on Sonnet. Otherwise Sonnet 5 handles it faster and cheaper.

Sonnet 5 costs $2 per million input and $10 output at its introductory rate (through 31 August 2026; $3/$15 after) and handles most day-to-day work without blinking. We default to Sonnet and escalate only when a task genuinely needs the extra reasoning. Our Sonnet vs Opus comparison goes deeper.

The honest rule we follow: if Sonnet solves it in two tries, keep using Sonnet. When the same prompt fails repeatedly or the job runs for hours untouched, the accuracy of the bigger model pays for its price.

Review: real strengths and honest limits

Strong, consistent instruction-following across long unattended sessions — but it's no longer the model to start on, and pointing it at trivial work burns money.

The strengths are real: it follows instructions consistently across long engineering sessions, carries context and style better than 4.7, and it's faster despite thinking harder. Simon Willison called it "a modest but tangible improvement," and that matches our experience — the comment-verbosity and tool-calling quirks of Opus 4.7 are largely gone.

Coding is where it separates from the pack. It's built for complex agentic work — planning a change, editing across files, running tests, and fixing what breaks — and it holds instructions steadily enough for unattended engineering. That reliability is why it shipped generally available in GitHub Copilot on launch day.

The limits are just as real. It is no longer the current Opus: Opus 5 is a step up at exactly the same price, so a new project choosing 4.8 is leaving capability on the table for free. At $5/$25 it's expensive for trivial tasks. Like every Claude model, it does not generate images — it can analyse them, but image creation isn't a feature. And because thinking is off by default here, teams sometimes benchmark it without thinking enabled and wonder why the reasoning gains didn't show up. Our Opus release history traces where this version sits in the arc of the line.

The quick version:

• Tops SWE-bench Pro at 69.2% for agentic coding
• Costs $5 in / $25 out per million tokens — same as Opus 5
• Superseded by Opus 5; keep it for pinned deployments and refusal fallbacks
• Reach for it only when Sonnet 5 falls short

Claude pricing at a glance

Opus-tier access needs Pro ($20/mo) or above; on the API you pay $5/$25 per million tokens whether you call Opus 5 or 4.8.

For the full breakdown of every plan, see our how much Claude costs guide, and Anthropic's official pricing page for current per-token rates. Ready to move up? Our Claude Opus 5 guide covers the two breaking changes to check before you switch the model string.