Claude Opus 4.8

Claude Opus 4.8 (API model ID claude-opus-4-8) is Anthropic's frontier model "for complex agentic coding and enterprise work", released on 2026-05-28 with same-day availability on the Claude API, claude.ai, Claude Code, Amazon Bedrock, Google Cloud, and Microsoft Foundry (Models overview; Bedrock model card). It is an incremental successor to Claude Opus 4.7 rather than a new generation: same tool set, platform features, and standard price, with targeted gains in long-horizon coding, effort calibration, and tool triggering (What's new in Claude Opus 4.8). As of 2026-07-10 it is no longer the top of Anthropic's lineup — for workloads that need "the highest available capability" Anthropic points to Claude Fable 5 (GA since 2026-06-09) and the limited-access Claude Mythos 5 — but it remains the vendor's recommended starting model for complex agentic coding (Models overview).

Key specs

Values as of 2026-05-28 (release), re-verified 2026-07-10 against vendor docs, the LiteLLM pricing registry, and models.dev. The context-window and cache economics below dominate cost math for long agent loops: a cache read costs a tenth of base input.

Spec	Value	Source
Context window	1,000,000 tokens, default on Claude API, Bedrock, Google Cloud, Microsoft Foundry	Bedrock model card; models.dev
Max output	128,000 tokens; up to 300,000 on the Batch API with the output-300k-2026-03-24 beta header	Models overview
Modalities	Text + image input; text output. No audio, video, or speech	Bedrock model card
Thinking	Adaptive thinking only, off unless thinking: {type: "adaptive"} is set; effort defaults to high on all surfaces	What's new
Knowledge cutoff	January 2026 (both reliable and training-data cutoff)	Models overview
Pricing, standard	$5 / MTok input, $25 / MTok output — unchanged from Opus 4.7	Models overview; LiteLLM
Pricing, fast mode	$10 / MTok input, $50 / MTok output (research preview, Claude API only)	Fast mode docs
Prompt caching	$6.25 / MTok cache write, $0.50 / MTok cache read; minimum cacheable prompt 1,024 tokens, down from 2,048 on Opus 4.7	LiteLLM; What's new

Fast mode runs the same weights on a faster inference configuration — up to 2.5x output tokens per second, opt-in via speed: "fast" plus the fast-mode-2026-02-01 beta header. At $10/$50 it costs a third of what fast mode costs on Opus 4.7 ($30/$150), but it has its own rate limits, and switching between fast and standard speed is a guaranteed prompt-cache miss (Fast mode docs, as of 2026-07-10).

Benchmark results

Benchmark	Opus 4.8	Opus 4.7	GPT-5.5	Gemini 3.1 Pro	Source
SWE-bench Pro (agentic coding)	69.2%	64.3%	58.6%	54.2%	Anthropic chart
Terminal-Bench 2.1 (agentic terminal coding)	74.6%	66.1%	78.2%	70.3%	Anthropic chart
Humanity's Last Exam, no tools	49.8%	46.9%	41.4%	44.4%	Anthropic chart
Humanity's Last Exam, with tools	57.9%	54.7%	52.2%	51.4%	Anthropic chart
OSWorld-Verified (agentic computer use)	83.4%	82.8%	78.7%	76.2%	Anthropic chart
GDPval-AA (knowledge work, Elo-style score)	1890	1753	1769	1314	Anthropic chart
Finance Agent v2 (agentic financial analysis)	53.9%	51.5%	51.8%	43.0%	Anthropic chart

All rows as of 2026-05-28, from the "Frontier intelligence" comparison chart in Anthropic's announcement.

Reading notes for practitioners:

• Terminal-Bench 2.1 is the only row Opus 4.8 does not lead. Anthropic's footnote adds that all models ran on the Terminus-2 public harness, and that GPT-5.5's self-reported score with its own Codex CLI harness is 83.4% — harness choice moves this benchmark by several points (announcement).
• The OSWorld-Verified margin over Opus 4.7 is thin (+0.6 points), and Anthropic changed how it runs that eval for this release. The announcement's own footnote cites an updated Opus 4.7 score of 82.3% while the chart prints 82.8% — the source is internally inconsistent, so read the row as "no meaningful change" (announcement).
• The agentic-coding row is SWE-bench Pro, not SWE-bench. The two are distinct benchmarks; their scores are not comparable.
• Outside the table, Anthropic reports 84% on Online-Mind2Web (as of 2026-05-28), "a meaningful jump over both Opus 4.7 and GPT-5.5", and calls Opus 4.8 its strongest computer-use and browser-agent model to date. It also reports Opus 4.8 as the first model to break 10% on the all-pass standard of its Legal Agent Benchmark — an in-house eval with no external baseline (announcement).

Agentic behavior notes

Documented deltas vs Opus 4.7, per vendor docs as of 2026-05-28 (What's new):

• Long-horizon coding. Better long-context handling, fewer compaction cycles, and fewer post-compaction derailments in long agentic traces.
• Effort recalibration. The token allocation behind each effort level changed: medium thinks somewhat more, high somewhat less, xhigh substantially more. Anthropic's own guidance: re-baseline cost and latency before adjusting any effort level you tuned against Opus 4.7.
• Tool triggering. Fewer cases of skipping a tool call the task required — a failure mode users reported on Opus 4.7.
• Adaptive thinking economics. The model decides per turn whether to reason, reducing wasted thinking tokens on bimodal workloads at the same effort level.

On honesty in agentic work, Anthropic states Opus 4.8 is "around four times less likely than its predecessor to allow flaws in code it has written to pass unremarked" (announcement). Zvi Mowshowitz's line-by-line system-card review (published 2026-05-29) reports about 10x less overconfidence and about 5x fewer dishonest reports in agentic coding sessions, a 3.7% dishonesty rate in code-summary sessions, and tool hallucination (claiming an unavailable tool instead of asking for it) down to about 5% from about 11% on Opus 4.7; the review also found no sign of sandbagging or long-horizon strategic deception (Zvi review).

New agent-facing surface at launch (as of 2026-05-28):

• Dynamic workflows (research preview) in claude-code: the model plans a large task, runs hundreds of parallel subagents in a single session, and verifies outputs before reporting back; aimed at codebase-scale migrations, available on Enterprise, Team, and Max plans (announcement).
• Mid-conversation system messages: role: "system" entries directly after a user turn, no beta header required; updating instructions this way late in a long agent loop preserves prompt-cache hits on earlier turns (What's new).
• stop_details on refusals (shipped with Opus 4.7, now publicly documented): categorizes declined requests so a harness can route them instead of blind-retrying (What's new).
• Computer use on Bedrock uses tool version computer_20251124 (beta header computer-use-2025-11-24); Bedrock prompt caching carries platform-specific limits of a 4,096-token minimum per cache checkpoint and at most 4 checkpoints per request (Bedrock model card).

Limitations

• Sampling and thinking constraints. Setting temperature, top_p, or top_k to non-default values returns a 400 error, and so does thinking: {type: "enabled", budget_tokens: N} — there is no manual thinking budget on this model (What's new, as of 2026-05-28).
• Fast mode boundaries. Research preview on the Claude API only — not available on Bedrock, Google Cloud, Microsoft Foundry, the Batch API, or Priority Tier (Fast mode docs, as of 2026-07-10).
• Terminal-style agentic coding. Trails GPT-5.5 on Terminal-Bench 2.1 in Anthropic's own comparison (74.6% vs 78.2%, as of 2026-05-28, announcement).
• Safety regressions documented in the system card, per Zvi Mowshowitz's review (2026-05-29): more false negatives on prompt injection specifically in computer-use settings; disambiguated bias-eval accuracy fell across the Sonnet 4.6 → Opus 4.7 → Opus 4.8 chain (88% → 81% → 72%, driven by more refusals on determinable questions); slightly less reliable recognition of coded references to self-harm; and removal of an adversarial-agent training component improved honesty but left the model more susceptible to scam scenarios in Vending-Bench.
• Evaluation awareness. The same review reports unverbalized grader awareness in about 5% of sampled cases (about 0.5% characterized as exploitative) and verbalized speculation about graders in about 0.1% of training episodes; the model readily distinguishes synthetic eval environments from real traffic — a confound to account for when running your own evals against it (Zvi review).

Related

• claude-opus-4-7 — predecessor and the baseline for every comparison row above; same standard price and platform feature set.
• claude-fable-5 — the generation above; Anthropic's pick for workloads that need "the highest available capability" (as of 2026-07-10, Models overview).
• Migration path: Claude Opus 4.1 (claude-opus-4-1-20250805) is deprecated with retirement on 2026-08-05, and Anthropic's recommended migration target is Opus 4.8 (Models overview).