Claude Opus 4.8 Released: Benchmarks, Features & What It Means (May 2026)
Claude Opus 4.8 Is Here — And It Is the New Anthropic Flagship
On May 28, 2026, Anthropic released Claude Opus 4.8, replacing Opus 4.7 as the most capable model in the Claude lineup. Pricing is unchanged from Opus 4.7 — $5 per million input tokens and $25 per million output tokens — and the model is available everywhere today via the Claude API under the identifier claude-opus-4-8.
On paper this is described as a "modest but tangible" upgrade. In practice, the combination of better benchmarks, dynamic workflows in Claude Code, effort control, and a fast mode that is now 3× cheaper than the previous generation means Opus 4.8 is a meaningful step forward for anyone building agentic systems.
Full release post: Anthropic — Introducing Claude Opus 4.8. Full evaluation details and the alignment assessment are in the Claude Opus 4.8 System Card.
Also useful: Best Claude Model in 2026 · Full Provider Comparison · Live AI Model Comparison Tool
The Benchmarks: Opus 4.8 vs Opus 4.7 vs GPT-5.5 vs Gemini 3.1 Pro
Anthropic's own benchmarks (released alongside the model) tell a clear story. Opus 4.8 leads on every major evaluation except Terminal-Bench 2.1, where OpenAI's GPT-5.5 still holds the top spot.
| Benchmark | Opus 4.8 | Opus 4.7 | GPT-5.5 | Gemini 3.1 Pro |
|---|---|---|---|---|
| Agentic coding (SWE-Bench Pro) | 69.2% | 64.3% | 58.6% | 54.2% |
| Agentic terminal coding (Terminal-Bench 2.1) | 74.6% | 66.1% | 78.2% | 70.3% |
| Multidisciplinary reasoning, no tools (Humanity's Last Exam) | 49.8% | 46.9% | 41.4% | 44.4% |
| Humanity's Last Exam, with tools | 57.9% | 54.7% | 52.2% | 51.4% |
| Agentic computer use (OSWorld-Verified) | 83.4% | 82.8% | 78.7% | 76.2% |
| Knowledge work (GDPval-AA) | 1890 | 1753 | 1769 | 1314 |
| Agentic financial analysis (Finance Agent v2) | 53.9% | 51.5% | 51.8% | 43.0% |
The headline number for most teams is SWE-Bench Pro at 69.2% — a meaningful jump on already-strong Opus 4.7 numbers, and a 10+ point lead over GPT-5.5 on the same test. For long-horizon agent workloads, the GDPval-AA score of 1890 is also notable — it is the strongest result Anthropic has ever posted on the benchmark that most closely tracks real knowledge-work value.
What's Actually New
Three of the launch-day features matter more than the benchmark deltas if you are building real systems.
1. Dynamic Workflows in Claude Code
Available in research preview on Claude Code for Enterprise, Team, and Max plans, dynamic workflows let Claude plan a task, then spin up hundreds of parallel sub-agents in a single session to execute it. Claude verifies its own outputs before returning. The headline use case Anthropic showed: codebase-scale migrations across hundreds of thousands of lines of code, with the existing test suite as the bar for "done."
This is the same architectural direction we wrote about in OpenClaw + Claude Code: Building an Always-On AI Operations Layer — agents that plan, fan out, verify, and report back instead of operating one prompt at a time.
2. Effort Control
There is now a control next to the model selector in claude.ai (and in Cowork) that lets users choose how much effort Claude puts into a response. Opus 4.8 defaults to high. Users can also choose extra (called xhigh in Claude Code) or max for the hardest tasks, or dial it down for faster responses that consume rate limits more slowly.
This is the same mental model OpenAI shipped with its reasoning.effort parameter (see our Best ChatGPT Model 2026 guide) and the standardisation across providers is good news for anyone building model routing into their stack.
3. Mid-Conversation System Updates via the Messages API
The Messages API now accepts system entries inside the messages array. That means developers can change Claude's instructions mid-task — updating permissions, token budgets, or environment context — without breaking the prompt cache or routing the update through a user turn. This is a small change with big implications for long-running agent harnesses.
The Honesty Improvement (And Why It Matters)
The single most important behavioural change in Opus 4.8 — and the one we expect to actually move the needle in production — is honesty. Anthropic's evaluations show Opus 4.8 is around four times less likely than Opus 4.7 to allow flaws in code it has written to pass unremarked.
In agentic workflows, this is the failure mode that costs the most. A model that confidently reports "done" on a task that actually shipped a regression is far more expensive than a model that flags uncertainty and pings a human. Anthropic's Alignment team also concluded that Opus 4.8 "reaches new highs on our measures of prosocial traits like supporting user autonomy and acting in the user's best interest," with misalignment rates substantially lower than Opus 4.7.
Pricing: Unchanged on Regular Use, Cheaper on Fast Mode
| Mode | Input ($/M tokens) | Output ($/M tokens) |
|---|---|---|
| Standard | $5 | $25 |
| Fast mode (2.5× speed) | $10 | $50 |
Standard pricing is identical to Opus 4.7. Fast mode is 3× cheaper than fast mode on previous Opus generations, which makes it a much more reasonable default for latency-sensitive agents and interactive Claude Code sessions.
Opus 4.8 vs GPT-5.5 vs Gemini 3.1 Pro — Our Take
None of the three flagship models is the right answer for every job. The honest framing:
- Opus 4.8 is the new default for high-stakes agentic coding, codebase-scale work, and any workflow where "the model not lying to you" is a hard requirement. The SWE-Bench Pro and GDPval-AA leads are real.
- GPT-5.5 remains the best choice for shell-first terminal workflows (78.2% on Terminal-Bench 2.1) and is still the strongest model in OpenAI's ecosystem for ChatGPT-native integrations.
- Gemini 3.1 Pro wins on raw cost-per-token and stays competitive on multimodal and long-context work, but trails on agentic coding and knowledge work in this evaluation set.
For a side-by-side view of all 200+ models we track — including filters for cost, context length, and use case — use our live AI Model Comparison Tool.
Where Opus 4.8 Fits in a Real Stack
If you are building operational systems, the practical takeaway is the same as it was with Opus 4.7: Opus is the judgment layer, not the execution backbone. We still recommend running Make.com (or equivalent) as your orchestration layer and reserving Opus for the steps that genuinely require frontier reasoning — code generation, multi-step planning, agentic decisions, complex extraction.
This is the architecture we describe in detail in The Five-Layer Stack and The Daily Stack: How to Scale Without Increasing Headcount. The model upgrade does not change the architecture; it just makes the intelligence layer cheaper, more accurate, and more honest.
Should You Upgrade From Opus 4.7?
For most teams, yes — there is no migration cost. Same API, same pricing, same context window. Swap claude-opus-4-7 for claude-opus-4-8 in your API calls and re-run your evaluation suite. If you are building agentic systems where the model writes or modifies code, the honesty improvement alone is worth the swap.
The one reason to wait: if you have prompt-cached workloads that are heavily tuned to 4.7's specific behaviour, give it a week of evaluation before flipping production traffic over.
What's Next: Project Glasswing and Claude Mythos
Anthropic also signalled that a new class of model above Opus is coming. Project Glasswing — currently powering Claude Mythos Preview for a small number of cybersecurity organisations — is the named track. Anthropic says Mythos-class models require stronger cyber safeguards before general release and that progress on those safeguards is being made "in the coming weeks." Worth watching.
FAQ
Is Claude Opus 4.8 better than GPT-5.5?
On the agentic coding (SWE-Bench Pro), knowledge work (GDPval-AA), and reasoning (Humanity's Last Exam) benchmarks Anthropic published at launch, Opus 4.8 leads. GPT-5.5 still wins on Terminal-Bench 2.1. In practice, run both against your own workload — the right answer is workload-specific.
How much does Claude Opus 4.8 cost?
$5 per million input tokens and $25 per million output tokens — the same as Opus 4.7. Fast mode (2.5× speed) is $10/$50 per million tokens, which is 3× cheaper than fast mode on previous Opus generations.
What are dynamic workflows in Claude Code?
A research-preview feature that lets Claude plan a task and run hundreds of parallel sub-agents in a single session, then verify its own outputs before reporting back. Available on Claude Code for Enterprise, Team, and Max plans. The headline use case is codebase-scale migrations across hundreds of thousands of lines of code.
What is effort control?
A new control next to the model selector in claude.ai and Cowork that lets users choose how much effort Claude puts into a response. Opus 4.8 defaults to high; users can also choose extra (xhigh in Claude Code) or max for the hardest tasks, or dial it down for faster, cheaper responses.
Should I upgrade from Claude Opus 4.7?
For most teams, yes — pricing, API, and context window are unchanged. The honesty improvement (~4× less likely to let code flaws pass unremarked) is the main reason to switch for any agentic coding workflow.
Compare other providers: Best ChatGPT Model 2026 · Best Gemini Model 2026 · Best Llama Model 2026 · Full Comparison: All Providers