Claude Opus 4.8: New Features, Benchmarks, and Pricing
Anthropic released Claude Opus 4.8 on 28 May. See verified benchmarks vs GPT-5.5 and Gemini 3.1 Pro, new pricing, and what to pin if you build on the API.
On this page
Anthropic released Claude Opus 4.8 on May 28, 2026, less than two months after Opus 4.7. The headline numbers are up across coding, reasoning, and alignment. The pricing surprise sits in fast mode, which is now three times cheaper than the previous generation. This guide breaks down what changed, how the new model stacks up against GPT-5.5 and Gemini 3.1 Pro, and what cost-conscious teams should do about it.
What Anthropic Shipped on May 28
Opus 4.8 looks like a point release, but it ships several material changes at once.
- Standard pricing held steady at $5 per million input tokens and $25 per million output tokens. That matches Opus 4.7.
- Fast mode is roughly three times cheaper. Pricing dropped from $30/$150 to $10/$50 per million tokens, while keeping the 2.5x speed uplift over standard.
- Effort modes are configurable per request, from Low through Max. The model defaults to High, which Anthropic judges the best balance of quality and latency.
- Dynamic Workflows entered research preview inside Claude Code. The runtime orchestrates up to 1,000 parallel subagents per task, capped at 16 concurrent.
- Honesty and alignment improved measurably. Anthropic positions Opus 4.8 as its most honest flagship to date, close to its specialised Claude Mythos Preview model.
- Distribution is wide on day one. The model is live on the Anthropic API, AWS Bedrock, Snowflake Cortex, and GitHub Copilot.
If you use Claude Code or Cowork, you do not need to act. New sessions auto-update to Opus 4.8. The upgrade question is really for teams pinning model versions on the API.
Opus 4.8 Benchmarks at a Glance
Across the benchmarks Anthropic and independent reviewers reported at launch, Opus 4.8 lifts every score that mattered for Opus 4.7. The biggest gains land in agentic coding.
Coding: SWE-Bench and Terminal-Bench
- SWE-bench Verified: 88.6%, up from 87.6% on Opus 4.7.
- SWE-bench Pro: 69.2%, up from 64.3%. This is the harder real-world repo task. The jump is the standout figure.
- Terminal-Bench 2.1: 74.6%, up from 66.1%. Shell-driven agents see a meaningful lift, though GPT-5.5 still leads this specific benchmark.
Reasoning: Humanity's Last Exam
Opus 4.8 scores 49.8% on Humanity's Last Exam without tools, and 57.9% with tools enabled. Both numbers put it ahead of GPT-5.5 and Gemini 3.1 Pro on the same evaluation. That matters for workloads that mix research, analysis, and long-chain reasoning.
Agentic Tasks and Tool Use
The agentic story is more nuanced. Opus 4.8 averages 80.1 across agentic benchmarks, just behind GPT-5.5 at 81.5. Where Opus 4.8 separates is coding-flavoured agentic work, averaging 76.4 against 58.6 for GPT-5.5. Anthropic also called out a new high on its Legal Agent Benchmark, and noted Opus 4.8 is the first model to break 10% overall on the all-pass standard.
Opus 4.8 vs GPT-5.5 vs Gemini 3.1 Pro
The three flagships stopped competing on raw capability some time ago. They now trade strengths. Here is the practical split.
Where Each Model Wins
- Claude Opus 4.8 is the pick for high-stakes code where correctness, edge cases, and reviewability matter more than raw speed. It is also the strongest of the three on multidisciplinary reasoning.
- GPT-5.5 still owns speed-critical agentic loops and terminal-driven tasks, leading Terminal-Bench 2.1 at 78.2%. If your agent shells out commands and reads stdout in a tight loop, this is the model to beat.
- Gemini 3.1 Pro retains the long-context crown with a 1M token window and the best multimodal pipeline. If your workload involves whole codebases, multi-document analysis, or images and PDFs as inputs, it is hard to displace.
Cost Per Benchmark Point
Per-token pricing sits inside a tight band across the three flagships. The cost-per-quality-point picture skews the decision.
| Model | Input per 1M | Output per 1M | SWE-Bench Pro | Coding Avg |
|---|---|---|---|---|
| Claude Opus 4.8 | $5.00 | $25.00 | 69.2% | 76.4 |
| GPT-5.5 | $5.00 | $30.00 | 58.6% | 58.6 |
| Gemini 3.1 Pro | $2.00 | $12.00 | 54.2% | Long-context tier |
On a typical 10K-token coding task, Opus 4.8 lands roughly 18% higher on SWE-bench Pro than GPT-5.5, at a 17% lower output token rate. Gemini 3.1 Pro is cheaper per token but trails on the coding benchmarks. For a full breakdown across providers, our side-by-side AI API comparison keeps the numbers current.
Pricing and the New 3x Cheaper Fast Mode
The most under-reported part of the release is fast mode. Anthropic held the standard tier at $5 and $25 per million tokens. It cut fast mode from $30/$150 to $10/$50, while keeping the 2.5x speed uplift. That makes fast mode genuinely competitive for latency-sensitive applications, instead of a premium emergency lever.
A rough rule of thumb after launch:
- Use standard mode for batch tasks, code generation, and anything where users tolerate a few seconds of wait.
- Use fast mode for real-time copilots, chat interfaces, and inline editor completions where every 500ms shows up in product metrics.
- Use Max effort with standard mode for one-shot evaluations, audits, and research where you want depth and do not care about response time.
Not sure which mix is right for your workload? Our AI pricing calculator will model the trade-off against your actual token volumes.
Honesty, Alignment, and Less Deception
Anthropic put alignment at the front of the announcement. The supporting numbers are concrete rather than vibes. Opus 4.8 is reported to be around four times less likely than Opus 4.7 to allow flaws in its own code to pass unflagged. Misaligned behaviour rates are described as substantially lower than 4.7, and similar to Claude Mythos Preview, Anthropic's best-aligned model to date.
In practice, that shows up as a model that more readily says "I am not sure". It flags uncertain assumptions in generated code, and pushes back on requests that previously produced a confident but wrong answer. For teams building developer tools, this should cut silent-failure rates in autonomous loops, which is where most production incidents on Opus 4.7 came from.
Effort Modes and Dynamic Workflows in Claude Code
Two new control surfaces shipped alongside the model.
Effort modes sit next to the model selector. Users on Claude.ai, Cowork, and Claude Code can pick Low through Max on a per-request basis. Low burns rate limits more slowly and replies faster. Max runs longer chains of thought, parallel verification, and self-critique before answering. Opus 4.8 defaults to High, calibrated to feel like an ordinary chat response while doing more thinking than the equivalent Opus 4.7 turn.
Dynamic Workflows is more ambitious. It is a research preview in Claude Code, needs version 2.1.154 or later, and is available on Max, Team, and Enterprise plans. The runtime lets Claude write a JavaScript orchestration script on the fly, then executes it to spin up subagents in parallel. Hard caps sit at 16 concurrent and 1,000 total per run.
The flagship use cases Anthropic showed include codebase audits, large refactors, and exhaustive test suite generation. Adversarial subagents inside the workflow try to refute findings before the result returns to the user. That feedback loop is where the four-times-fewer-unflagged-flaws claim partly comes from.
What Changes for You (Claude Code, Cowork, and the API)
Where the upgrade lands depends on how you consume Claude.
- Claude Code and Cowork users get Opus 4.8 automatically on new sessions, provided Claude Code is on v2.1.154 or newer. No action is required, and there is no easy way to roll back to 4.7 inside the CLI.
- Claude.ai users see Opus 4.8 as the default flagship in the model picker, with effort controls exposed next to it.
- API and SDK builders control their own destiny. The model ID for the new release is distinct from Opus 4.7. Anything pinned to 4.7 keeps running 4.7 until you change it. That pinning decision is the real upgrade question for production teams.
For most API teams, migration risk is low because prompts written for Opus 4.7 generally port without changes. The behaviour deltas worth testing are the new honesty floor, which can change refusal patterns, and any prompt that previously relied on the model not pushing back on ambiguous instructions. If you maintain a model strategy across multiple providers, our AI API comparison and migration guides can shortcut the diff work.
How to Access Claude Opus 4.8
- Claude.ai: Default flagship as of May 28, on free and paid tiers.
- Cowork: Auto-updated on new sessions.
- Claude Code v2.1.154 or later: Auto-updated on new sessions across CLI, Desktop, and the VS Code extension.
- Anthropic API: Pin the Opus 4.8 model identifier.
- AWS Bedrock: Generally available from launch day.
- Snowflake Cortex AI: Generally available from launch day.
- GitHub Copilot: Generally available, selectable from the model picker.
For up-to-date token rates across every provider, see our Claude pricing breakdown and cheapest AI APIs ranked pages.
Fun Fact: Smarter and Cheaper in the Same Release
Opus 4.8 is one of the rare flagship releases where the model got noticeably stronger and noticeably cheaper at the same time. Fast mode pricing dropped from $30/$150 to $10/$50 per million tokens, a three times cut, in the same release where SWE-bench Pro jumped 5 points (64.3% to 69.2%) and Terminal-Bench 2.1 jumped 8.5 points (66.1% to 74.6%). Releases usually pick one of those two. Anthropic delivered both in a single drop.
The Bottom Line
Claude Opus 4.8 is a strong upgrade if your workload leans toward careful code, multi-step reasoning, or autonomous agents that need to flag their own uncertainty. GPT-5.5 still has the edge for speed-critical terminal agents. Gemini 3.1 Pro still wins for long-context and multimodal jobs. The fast mode price cut is the under-reported story, and changes the cost calculus for real-time products that previously priced themselves off the standard tier.
If you build on Claude Code or Cowork, you are already on 4.8. If you build on the API, the question is whether to update your pinned model ID. For most teams, the right answer is yes, after a short regression pass.
