Claude Opus 4.8: A Modest Bump That Quietly Tops the Leaderboard

TL;DR: Anthropic released Claude Opus 4.8 (claude-opus-4-8) on May 28, 2026 — just 41 days after Opus 4.7. Pricing is unchanged at $5/$25 per million tokens, context stays at 1M (128K output). The benchmark gains are real but incremental: SWE-Bench Pro 69.2% (from 64.3%), GDPval-AA 1890 Elo (from 1753), USAMO 2026 up 27 points to 96.7%. The headline features are product-side — Claude Code "dynamic workflows" that orchestrate up to 1,000 subagents, an effort control on claude.ai, and a fast mode that's now 3x cheaper. Simon Willison called it "a modest but tangible improvement." That's accurate. Here's what actually matters for builders.

What Is Opus 4.8?

Opus 4.8 is Anthropic's new flagship, replacing Opus 4.7 at the top of the Claude lineup. The model ID is claude-opus-4-8. The cadence is the story before the benchmarks are: 4.7 shipped in mid-April, 4.8 arrived May 28. Forty-one days. Anthropic is now iterating on the frontier roughly monthly, and the version number is doing less and less work — this is a point release that happens to lead several leaderboards.

Three things define this release:

1. It's a re-tune, not a re-architecture. Anthropic's own framing is "sharper judgment, more honesty about its progress, and the ability to work independently for longer." The API surface is identical to 4.7 — same adaptive thinking, same effort levels, same removed sampling parameters. If your code runs on 4.7, it runs on 4.8 with a string swap.

2. The gains concentrate in agentic and knowledge work. Coding, long-horizon autonomy, and the GDPval-AA economic-value benchmark move the most. Pure context-length and modality coverage are unchanged.

3. The product features are the real news. Dynamic workflows and a cheaper, faster fast mode change what you can ship more than the benchmark deltas do.

What Are the Verified Benchmarks?

A note on sourcing, because it matters here. The numbers below come from Anthropic's official announcement, cross-referenced against Artificial Analysis, llm-stats, and independent coverage from Vellum and The New Stack. Where a figure is vendor-reported and not yet independently reproduced, I mark it vendor-claimed. Treat the coding and knowledge-work numbers as well-corroborated; treat single-source claims with appropriate skepticism.

Two of these deserve more than a row in a table.

GDPval-AA at 1890 is the standout. It measures real, economically valuable knowledge work across professional domains, scored as an Elo. Opus 4.8 doesn't just lead — independent coverage puts its margin at roughly 576 points over Gemini 3.1 Pro on this benchmark. That's the widest spread on the board, and it aligns with the agentic-execution story Anthropic is telling. The +137 jump over 4.7 is the largest single-version GDPval-AA improvement Anthropic has reported.

USAMO 2026 going from 69.3% to 96.7% is a 27-point leap on rigorous, proof-style mathematics. This is the kind of number that's easy to over-read — olympiad benchmarks are narrow — but a near-saturation score on USAMO is a genuine signal about multi-step reasoning rigor, not a rounding artifact.

The one Opus 4.8 loses: Terminal-Bench 2.1, where GPT-5.5 still edges it (78.2% vs 74.6%). On terminal-heavy CLI automation, OpenAI keeps the crown — though Opus 4.8's +8.5 jump over 4.7 narrows the gap considerably.

How Does It Compare to GPT-5.5, Gemini 3.5, and the Field?

Where Opus 4.8 sits against the June 2026 frontier:

A few honest caveats. The Grok 4.3 comparison that gets asked for isn't well-covered in the launch-window sources I could verify, so I'm leaving it out rather than inventing rows. Gemini 3.5 Flash is the genuine surprise here — it posts a 76.2% on Terminal-Bench 2.1, ahead of Opus 4.8 on that single benchmark, at roughly a quarter of the cost and several times the speed. If your workload is high-volume agentic execution where the marginal task is cheap and you can tolerate the occasional miss, Flash is a serious option. Opus 4.8 wins when silent errors are expensive and the work is genuinely hard.

The positioning that holds up across sources: Opus 4.8 is the intelligence leader, GPT-5.5 is the terminal/computer-use workhorse, and Gemini 3.5 Flash is the budget-and-speed champion. For a fuller cross-model breakdown, see the FrankX models tracker.

What's the Pricing?

Pricing is the easy part: it didn't move. Standard Opus 4.8 is $5 input / $25 output per million tokens — same as Opus 4.7 and 4.6. The 1M context window carries no long-context premium. You get a better model at the same price, which is the quiet through-line of every Opus release since the 67% cut at 4.6.

The new wrinkle is fast mode. Opus 4.8 fast mode runs at roughly 2.5x the speed of standard for $10/$50 per million — double the per-token cost, but Anthropic notes it's three times cheaper than fast mode was on previous Claude models. For latency-sensitive interactive products (coding assistants, live chat), that's a meaningfully better speed-per-dollar curve than fast mode used to offer. For batch and background work, standard mode remains the right call.

What Changed vs Opus 4.6 and 4.7?

If you're coming from Opus 4.7, almost nothing changes in your code. There are no new breaking changes — the request surface is identical. The deltas are behavioral and product-side:

If you're jumping from Opus 4.6, you inherit the 4.7 breaking changes first: budget_tokens returns a 400 (use thinking: {type: "adaptive"}), the sampling parameters temperature / top_p / top_k are removed (steer with prompting), and last-assistant-turn prefills 400 (use output_config.format). Thinking content is also omitted by default — set thinking: {type: "adaptive", display: "summarized"} if you surface reasoning to users.

The behavioral re-tune is worth a prompt review. Opus 4.8 narrates more than 4.7 by default and is more deliberate — on minor decisions it tends to pause and ask rather than just proceeding. If you want 4.7-style terseness and autonomy back, add an explicit silence-default and a "for minor choices, pick and note rather than ask" instruction. It's also more conservative about reaching for search, subagents, and file-based memory, so move the "when to use this" guidance into each tool's own description, not just the system prompt.

What Are Dynamic Workflows?

This is the feature that earns the release. In Claude Code, a dynamic workflow is a JavaScript script — which Claude writes for you from a task description — that orchestrates subagents at scale. A runtime executes it in the background while your session stays responsive. The limits, per Anthropic and MarkTechPost's coverage: up to 16 concurrent agents and 1,000 agents total per run.

The concrete claim: Claude Code on Opus 4.8 can carry out codebase-scale migrations across hundreds of thousands of lines — from kickoff to merge — using the existing test suite as its bar. Plan, fan out across files, run the migration in parallel, verify against tests, open the PR.

Underneath it sits a real API change: the Messages API now supports real-time updates to the messages array while an agent is active. You can change instructions, adjust permissions, modify token limits, or update context mid-task without disrupting prompt caching or forcing a new user turn. Concretely, you place a {"role": "system", ...} entry partway through the messages array (beta header mid-conversation-system-2026-04-07) to inject operator instructions without invalidating the cached prefix ahead of the whole conversation. That's the primitive that makes long-running, steerable agents practical — and it's available from Opus 4.7 onward, not 4.8-exclusive.

The honest framing: dynamic workflows are a Claude Code orchestration layer, not a new model capability. But "the model can now reliably drive 1,000 subagents through a 200K-line migration" is a more useful sentence than any single benchmark on the board.

What Does It Mean for Builders?

For developers

The migration is trivial and the upside is free. Swap claude-opus-4-7 for claude-opus-4-8, run your evals, ship. No breaking changes, same price, better coding and reasoning. The only thing to actually do is re-tune prompts for the behavioral shifts — narration and ask-rate — and that's optional polish, not a blocker.

{
  "model": "claude-opus-4-8",
  "thinking": { "type": "adaptive" },
  "output_config": { "effort": "high" }
}

One nuance on effort: on 4.8, don't reflexively reach for xhigh or max. The intelligence ceiling is higher, so start at high and sweep medium/high/xhigh on your own eval set. Higher effort up front often reduces total turn count and cost on agentic work — but the relationship isn't monotonic, and for some routes medium lands equally good results faster.

For agentic and long-horizon work

This is where 4.8 earns its keep. Give it the full task specification in one well-specified first turn and run at high or xhigh. The long-horizon coherence comes partly from reasoning more at each step, so a clear up-front goal plus generous effort produces more efficient and more accurate output than feeding the task piecemeal across turns. If you're on Claude's Managed Agents, state "done" as a gradeable rubric via an Outcome and let the harness iterate.

For content and knowledge work

The GDPval-AA and knowledge-work gains are the practical headline for creators. Opus 4.8's prose is warmer and less hedged than 4.7's by default — which means the style prompts you wrote to counter 4.7's clipped directness may now overcorrect. Re-baseline them. The 128K output ceiling (unchanged) still means complete long-form drafts in a single generation pass, and the 1M context still holds an entire content archive in one session.

For cost-conscious routing

The arrival of fast mode changes the routing math for interactive products. For a coding assistant or live agent where latency is user-visible, Opus 4.8 fast mode at $10/$50 and 2.5x speed may now beat dropping down to Sonnet — you keep Opus-tier intelligence and pay for speed only where it's felt. For everything background, standard Opus 4.8 at $5/$25 remains the default. And for genuinely high-volume, error-tolerant fan-out, Gemini 3.5 Flash is the honest budget alternative. This is the same routing discipline the broader Microsoft MAI frontier models coverage points at: match the model to the task's cost-of-error, not to the leaderboard.

Are There Breaking Changes?

For Opus 4.7 users: no. The request surface is identical. This is the cleanest Opus migration in recent memory — a model-string swap.

For Opus 4.6 (or older) users, you inherit the 4.7 breaking changes:

None of these are new to 4.8 — they all landed with 4.7. If you already made the 4.7 jump, you're done.

FAQ

Is Opus 4.8 better than GPT-5.5?

On most benchmarks, yes — Opus 4.8 leads SWE-Bench Verified (88.6% vs ~82%), SWE-Bench Pro (69.2% vs ~58.6%), and the GDPval-AA knowledge-work benchmark by a wide margin. GPT-5.5 keeps the lead on Terminal-Bench 2.1 (78.2% vs 74.6%), so for terminal-heavy CLI automation and computer-use workflows, GPT-5.5 still edges it. For coding, analysis, and knowledge work, Opus 4.8 is the stronger model at a comparable price.

How much does Opus 4.8 cost?

$5 per million input tokens, $25 per million output tokens — unchanged from Opus 4.7 and 4.6. The 1M context window carries no long-context premium. Fast mode runs at $10/$50 per million for roughly 2.5x the speed, which Anthropic says is three times cheaper than fast mode on previous Claude models.

What's the context window and max output?

1M input tokens and 128K output tokens — identical to Opus 4.7 and 4.6. Modalities are text, vision, and code. No change on this axis in 4.8.

Should I upgrade from Opus 4.7?

Yes, and it's low-risk. There are no breaking changes — swap claude-opus-4-7 for claude-opus-4-8, run your evals, and you get better coding and reasoning at the same price. Budget an hour to re-tune prompts for the behavioral shifts (more narration, higher ask-rate, more conservative tool use), but that's polish, not a requirement.

What are dynamic workflows and what's the subagent limit?

Dynamic workflows are a Claude Code feature where Claude writes a JavaScript orchestration script from your task description, then runs it in the background to coordinate subagents at scale — up to 16 concurrent and 1,000 total per run. The flagship use case is codebase-scale migrations across hundreds of thousands of lines, verified against the existing test suite, from kickoff to merge.

Which benchmark numbers are verified vs vendor-claimed?

The coding (SWE-Bench Verified/Pro, Terminal-Bench) and knowledge-work (GDPval-AA) figures are well-corroborated across Anthropic's announcement, Artificial Analysis, llm-stats, and independent outlets. GPQA Diamond (93.6%) and the USAMO 2026 jump (96.7%) are reported but lean more on Anthropic's own evals — treat them as vendor-claimed until third parties reproduce them. I omitted any Grok 4.3 comparison because I couldn't verify it from launch-window sources rather than guess.

Analysis by Frank — former Oracle AI architect who helped build Oracle's AI Center of Excellence, now building agentic systems independently and making music with AI. Published June 5, 2026 with benchmarks validated against Anthropic's official documentation, Artificial Analysis, llm-stats, and independent coverage. Vendor-claimed figures are marked as such.