Claude Opus 4.8

Overview

Claude Opus 4.8 is Anthropic's most capable Opus-tier model, released on May 28, 2026. It is the recommended starting point for complex agentic coding and enterprise work — Claude Fable 5 sits above it for workloads that need the absolute highest capability and can absorb double the price.

Opus 4.8 is a refinement release rather than a rearchitecture. It keeps the exact request surface of Opus 4.7 while sharpening judgment on agentic tasks, improving honesty, and cutting tool-calling overhead. Anthropic positions it as more reliable and less prone to the comment-verbosity and tool-calling quirks seen in 4.7.

For developers already on Opus 4.7, migration is a model-ID swap: claude-opus-4-8. There are no new breaking changes.

Capabilities

• Long-horizon agentic execution: State-of-the-art on long, autonomous runs — complex refactors and overnight coding sessions that complete without human correction. Anthropic recommends giving the full task specification up front in a single well-specified turn and running at high or xhigh effort.
• Sharper agentic judgment: Approximately 4× less likely than its predecessor to let a code flaw pass unremarked.
• Honesty: More likely to flag uncertainty and avoid unsupported claims — a measurable shift, not just a tone change.
• Tool-calling efficiency: Reaches the same outcome in fewer steps, with improved token efficiency.
• Computer use and browsing: An early tester quoted in the launch post reports 84% on Online-Mind2Web, ahead of both Opus 4.7 and GPT-5.5. Anthropic itself publishes no Online-Mind2Web score — the benchmark appears nowhere in the system card.
• Legal reasoning: Harvey AI calls Opus 4.8 the first model to break 10% overall on the Legal Agent Benchmark's all-pass standard. Anthropic's own system card measures 9.62% on that standard; both figures are reported below.
• Writing voice: Anthropic describes Opus 4.8 as tending "toward a direct, opinionated style with minimal validation-forward phrasing and sparing emoji use," and warns that prose style on long-form writing may shift — if your product relies on a specific voice, re-evaluate your style prompts.

Technical Specifications

• API model ID: claude-opus-4-8
• Context window: 1M tokens, at standard API pricing (no long-context premium)
• Max output: 128K tokens per request. On the Message Batches API, up to 300K output tokens via the output-300k-2026-03-24 beta header.
• Thinking: Adaptive thinking — thinking: {"type": "adaptive"}. Off when the parameter is omitted.
• Effort levels: low, medium, high (default), xhigh, max
• Training data cutoff: January 2026 (reliable knowledge cutoff: January 2026)
• Tokenizer: The tokenizer introduced with Opus 4.7. The same text produces roughly 30% more tokens than on Opus 4.6 and older models.
• Architecture: Transformer-based, developed under Anthropic's Constitutional AI approach.

Supported features

Effort, adaptive thinking, task budgets (beta), compaction (beta), context editing (beta), the memory tool, code execution, programmatic tool calling, tool search, high-resolution vision (up to 2576px on the long edge), prompt caching, Message Batches, structured outputs, and mid-conversation system messages.

API Behavior

Opus 4.8 shares Opus 4.7's request surface. If you are arriving from Opus 4.6 or earlier, four things return a 400 error:

• thinking: {"type": "enabled", "budget_tokens": N} — removed. Use thinking: {"type": "adaptive"} and control depth with output_config.effort.
• temperature, top_p, top_k — removed. Steer behavior with prompting.
• Last-assistant-turn prefills — removed. Use structured outputs (output_config.format) or a system-prompt instruction.
• thinking.display — defaults to "omitted", so thinking blocks arrive with an empty thinking field. Pass "summarized" if you surface reasoning to users.

Mid-conversation system messages

Unique to Opus 4.8 among current Claude models. Append {"role": "system", "content": "..."} to the messages array to deliver an operator instruction partway through a session without rewriting the top-level system prompt and invalidating the cached prefix. No beta header is required. Phrase these as context, not commands — override-style language works against the model's training.

Behavioral Notes

Opus 4.8 is noticeably more deliberate than 4.7, and the differences are prompt-tunable:

• It narrates more. Expect more text between tool calls and longer end-of-task wrap-ups. Scaffolding like "after every 3 tool calls, summarize progress" is now redundant — remove it. If a coding agent is too chatty, add an explicit silence-default instruction.
• It favors reasoning over tool calls. Anthropic states this "produces better results in most cases," and that effort is the lever: high or xhigh "show substantially more tool usage in agentic search and coding." If the model is skipping a tool you want used, describe in the prompt why and when to call it.
• It spawns fewer subagents by default. Steerable through prompting — give explicit guidance on when subagents are desirable.
• Tool triggering improved over 4.7. Anthropic notes the model "is less likely to skip a tool call the task required, an issue some users reported on Claude Opus 4.7."
• Effort is a dimension to test. Anthropic's guidance is to start with xhigh for coding and agentic use cases, use high for most other intelligence-sensitive workloads, and step down to medium or low only once your evals show the lower level holds quality.

Use Cases

• Autonomous software development: Overnight refactors, large migrations, and multi-file feature work driven from a single well-specified brief.
• Enterprise knowledge work: Financial analysis, legal document review, and research synthesis where consistency across a long document set matters.
• Code review and debugging: High-recall bug finding with clearer explanations. Note that instructions like "only report high-severity issues" are followed literally and can depress measured recall — report everything and filter downstream.
• Agent orchestration: Coordinating subagents and long-running tool loops, with task budgets to pace cumulative spend.
• Computer use: Browser and desktop automation. Anthropic's own computer-use figures are OSWorld and ScreenSpot-Pro; the widely quoted 84% Online-Mind2Web score comes from a tester, not from Anthropic.
• High-throughput interactive products: Fast mode delivers up to 2.5× output tokens per second for latency-sensitive surfaces.

Performance / Benchmarks

Anthropic's launch post publishes its Opus 4.8 comparison as an image, so no scores can be read out of the post itself. The figures below are transcribed from Table 8.1.A, "Capability evaluation summary," in the Claude Opus 4.8 System Card (May 28, 2026), which publishes the same numbers as text.

The caption states the setting that applies to every row unless otherwise noted: "adaptive thinking at max effort, default sampling settings (temperature, top_p), averaged over 5 trials." These are not default-effort scores — effort on Opus 4.8 defaults to high.

Four rows do not share the standard configuration:

• Terminal-Bench 2.1 was run at high effort, not max, on the Terminus-2 harness: "On Harbor's leaderboard, using the Terminus-2 harness, Opus 4.8 achieved 74.6% mean reward, averaged over 5 attempts for each one of the 89 unique tasks." Anthropic later changed harnesses — the Claude Fable 5 system card restates Opus 4.8 at 82.7% on the same benchmark under mini-SWE-agent. Same model, different harness; the two are not comparable. A third figure circulates: Z.ai's GLM-5.2 model card lists Opus 4.8 at 85 on Terminal-Bench 2.1 (Terminus-2), which Anthropic's own Terminus-2 result contradicts.
• GPQA Diamond is averaged over 25 trials rather than 5.
• Finance Agent v2 was measured by Vals AI, and GDPval-AA by Artificial Analysis. Neither is an Anthropic-run evaluation.

Legal Agent Benchmark

The Legal Agent Benchmark is Harvey AI's open-source suite of 1,200+ tasks, and its standard scoring counts a task as a success only if every rubric criterion passes. Anthropic's system card reports Opus 4.8 at a 9.62% all-pass rate and an 89.01% mean criterion-pass rate, tested against 1,235 problems at adaptive thinking / max effort, averaged over 5 trials.

Harvey's Niko Grupen, quoted in the launch post, calls Opus 4.8 "the first model to break 10% overall on the all-pass standard." That is Harvey's number; Anthropic's own measurement is 9.62%.

Figures that are testimonials, not measurements

The launch post's most-quoted numbers sit inside an early-tester quote block, not in Anthropic's evaluations:

• "84% on Online-Mind2Web" comes from a tester quote (Miguel Gonzalez, Tech Lead). Online-Mind2Web appears nowhere in the Opus 4.8 system card, and Anthropic publishes no score for it.
• Super-Agent ("the only model to complete every case end-to-end") and CursorBench ("exceeds prior Opus models across every effort level") are likewise tester statements, and neither carries a number.

One claim in the post is Anthropic's own, though it is not a benchmark score: "our evaluations ... 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."

Artificial Analysis

Opus 4.8's Artificial Analysis page reads: "Claude Opus 4.8 (Adaptive Reasoning, Max Effort) scores 56 on the Artificial Analysis Intelligence Index." The page labels that index v4.1. Scores are not comparable across index versions, so a v4.1 figure cannot be set against one from an earlier revision.

How to read these numbers

Every figure above is vendor-reported: Anthropic chose the benchmarks, the harnesses, and the effort level, and ran the evaluations itself except where a third party is named. Treat them as an upper bound and validate on your own evaluation set. Anthropic's own system card warns that "scores are not comparable across differently-configured harnesses" — and on Anthropic's own reporting, Terminal-Bench 2.1 alone moves from 74.6 to 82.7 depending on which harness is used.

Anthropic did not state an ASL designation for Opus 4.8 in the launch announcement; the system card dated May 28, 2026 carries the full safety evaluation.

Limitations

• Cost: $5/$25 per million tokens places it well above the Sonnet and Haiku tiers. Fast mode doubles that.
• Latency: Anthropic classifies its comparative latency as "moderate" — slower than Sonnet 5 and Haiku 4.5, faster than Fable 5.
• Verbose reasoning when thinking is disabled: With thinking: {"type": "disabled"}, Opus 4.8 occasionally writes long reasoning into the visible response. Leaving adaptive thinking on is the simplest fix.
• Reduced parameter surface: No temperature/top_p/top_k, no budget_tokens, no assistant prefill.
• Tokenizer shift: The same text costs roughly 30% more tokens than on Opus 4.6 and older. Re-baseline max_tokens, compaction triggers, and cost dashboards rather than applying a blanket multiplier.
• No internet access: Cannot browse unless integrated through tool use or the server-side web search tool.

Pricing & Access

API Pricing

Available on the Claude API, Claude Platform on AWS, Amazon Bedrock, Google Cloud, and Microsoft Foundry:

• Input: $5 / MTok
• Output: $25 / MTok
• Fast mode: $10 / MTok input, $50 / MTok output — described by Anthropic as 3× cheaper than fast mode on previous models
• Prompt caching: cache reads at roughly 10% of base input price; minimum cacheable prefix 1,024 tokens — lower than the 4,096 required on Opus 4.6 and 4.5
• Batch API: 50% discount

Fast mode is a research preview and is not available with the Batch API, Priority Tier, Claude Platform on AWS, or third-party platforms.

Consumer Plans

Claude's consumer tiers are Pro ($20/month) and Max (from $100/month), alongside a Free tier. Anthropic does not publish which model backs each tier, so we do not claim Opus 4.8 is reachable on any particular plan — check claude.com/pricing for current entitlements.

Ecosystem & Tools

Official SDKs

• Python SDK — full-featured with async support
• TypeScript SDK — type-safe for Node.js and browsers
• REST API — direct HTTP access from any language

SDKs also ship for Java, Go, Ruby, C#, and PHP, along with the ant command-line interface.

Developer Tools

• Anthropic Console — playground and prompt development
• Claude Code — agentic coding in the terminal, IDE, and web. On Opus 4.8 the effort parameter defaults to high on all surfaces, Claude Code included; set it explicitly to use another level
• Claude.ai — browser-based chat

Community & Resources

• Introducing Claude Opus 4.8 — official announcement
• Claude Opus 4.8 System Card — safety and capability evaluation, May 28, 2026
• Models Overview — specs for all current Claude models
• Migration Guide — upgrading from Opus 4.7 and earlier
• Pricing Page