Claude Opus 5 Is Out: Migrate from Opus 4.8 Now

Anthropic shipped Claude Opus 5 on July 24, outperforming GPT-5.6 Sol on agentic coding with a 43.3% score on Frontier-Bench v0.1 versus Sol's 34.4%, but the upgrade introduces two silent breaking changes: omitting the thinking field now consumes thinking tokens by default, sharing the max_tokens budget and causing truncated responses, and disabling thinking at effort levels xhigh or max returns HTTP 400. The fix requires increasing max_tokens by 2,000-5,000 tokens and either enabling thinking or dropping effort to high or below, while pricing remains $5 input / $25 output per million tokens with a new fast mode at $10/$50 per million tokens for 2.5x speed.

Anthropic shipped Claude Opus 5 on July 24, and if you’re running Opus 4.8 in production, two silent breaking changes are waiting for you. The headline numbers are worth celebrating — Opus 5 now outperforms GPT-5.6 Sol on agentic coding https://www.anthropic.com/news/claude-opus-5 and beats Anthropic’s own Fable 5 flagship on cost-adjusted benchmarks — but the breaking changes will hit you before you get to celebrate. Here’s what changed, what breaks, and how to fix it. The Breaking Change That Will Actually Bite You On Opus 4.8, omitting the thinking field from your API request meant no thinking tokens were consumed. Clean request, predictable token budget. On Opus 5, the same request now thinks by default. Your max tokens limit — which previously covered only the response — now covers thinking tokens plus response text. The failure mode is insidious: the request returns HTTP 200, stop reason comes back as "max tokens" , and your response is truncated mid-sentence. No error, no warning. This is the kind of bug that makes it into production because your test suite passes and your monitoring shows 200s. The fix is a two-step audit: Step 1: Update your model string Before model = "claude-opus-4-8" After model = "claude-opus-5" Step 2: Increase max tokens — thinking now shares this budget Rule of thumb: add 2,000-5,000 tokens to your existing limits max tokens = 16000 was 8000 on Opus 4.8 The second half of this breaking change: you cannot disable thinking at effort: "xhigh" or effort: "max" . Doing so returns HTTP 400. If your code disables thinking and runs at high effort levels, you need to pick one: enable thinking, or drop effort to "high" or below. This returns HTTP 400 on Opus 5: { "thinking": {"type": "disabled"}, "effort": "xhigh" ERROR } Fix option A: enable thinking { "thinking": {"type": "enabled"}, "effort": "xhigh" } Fix option B: drop effort { "thinking": {"type": "disabled"}, "effort": "high" OK } The Benchmark Story: Cheapest Path to Near-Frontier Here is what Opus 5 actually scores against the models it’s most likely replacing or competing with: | Benchmark | Opus 5 | Fable 5 | GPT-5.6 Sol | Opus 4.8 | |---|---|---|---|---| | Frontier-Bench v0.1 agentic coding | 43.3% | 33.7% | 34.4% | 18.7% | | ARC-AGI-3 novel reasoning | 30.2% | — | 7.8% | — | | SWE-bench Pro code repair | 79.2% | 80.0% | — | 69.2% | | GDPval-AA v2 knowledge work | 1,861 | — | 1,736 | — | The Frontier-Bench number is the one that matters most for teams running coding agents. Opus 5’s 43.3% against Sol’s 34.4% is a 26% relative lead on multi-step, tool-using coding work. Third-party analysis confirms Opus 5 leads 9 of 12 benchmarks https://codersera.com/blog/claude-opus-5-vs-gpt-5-6-2026/ against GPT-5.6 Sol. Fable 5 — Anthropic’s $15/$75 flagship — scores 33.7% on the same benchmark and costs three times as much per token. On ARC-AGI-3 https://arcprize.org/results/anthropic-claude-opus-5 , which tests novel reasoning on problems the model hasn’t seen before, Opus 5 scores roughly 3x higher than GPT-5.6 Sol. That’s not a marginal benchmark improvement; it’s a different tier of performance. The knowledge cutoff is also worth noting: Opus 5 carries a May 2026 cutoff, the freshest in the Claude lineup, compared to January 2026 for both Fable 5 and Opus 4.8. Pricing: Same Cost, New Fast Mode Standard pricing stays at $5 input / $25 output per million tokens — identical to Opus 4.8. The price-to-performance improvement is real: you’re getting significantly better results for the same bill. Fast mode is new: $10/$50 per million tokens double the cost for roughly 2.5x the speed. The math only works if latency is the bottleneck — interactive agents, user-facing tools where a two-second difference matters. For background jobs, batch processing, or overnight pipelines, stick with standard. Fast mode is also first-party API only for now; Bedrock and GCP users don’t have access as of July 27. The Effort Dial: A Quick Reference Opus 5 ships with five effort levels that control how deeply the model thinks before responding. Most teams won’t need to touch this, but here’s the decision tree: low — Skip thinking entirely. Batch classification, template filling, anything with a predictable format. medium — Light reasoning. Summaries, routine Q&A, everyday drafting. high — The default. Complex reasoning, standard coding tasks. Start here. xhigh — Extended thinking for long-horizon agentic work. Thinking must be enabled at this level. max — Maximum reasoning depth. Reserve for evals, critical production decisions, and tasks where quality trumps cost. The practical guidance: start at high , run your evals, and adjust from there. Stepping down to medium or low where quality holds saves real money at scale. Stepping up to xhigh or max for your most demanding agent tasks is where Opus 5’s gains over GPT-5.6 Sol are most pronounced. Migration Checklist If you’re on Opus 4.8 and ready to migrate, here’s the full checklist: - Change model string from claude-opus-4-8 to claude-opus-5 - Increase max tokens on every request — add at least 2,000-5,000 tokens to existing limits - Audit any code that sets thinking: disabled with effort: xhigh or max — these will return HTTP 400 - Run your existing test suite and check for truncated responses stop reason: max tokens - Claude Code users on Max plan: you’re already on Opus 5 — the model alias updated automatically on July 24 Claude Code users can run /claude-api migrate to get guided migration steps in the editor. This requires Claude Code v2.1.219 or later https://code.claude.com/docs/en/changelog . Should You Upgrade Now? If your workload is agentic coding, long-horizon reasoning, or computer use — yes, upgrade now. Opus 5 is a straightforward improvement over Opus 4.8 at the same price, and it beats GPT-5.6 Sol on the benchmarks that matter most for these workloads. The breaking changes are real and require a short audit pass, but budget an hour for testing on any production system before flipping the model string. If you’re running high-volume tasks with thinking disabled and no complex agent work, the migration is low-risk but the gains are smaller. Still worth doing — the fresher knowledge cutoff alone is a free upgrade. Fable 5 users: no compelling reason to switch yet. Opus 5 competes on cost-adjusted benchmarks but Fable 5 still edges it on SWE-bench Pro and gives you the familiar frontier-grade headroom for your most demanding tasks. The interesting case is teams currently paying $15/$75 for Fable 5 who could move most workloads to Opus 5 at one-third the cost — the benchmark gap on real coding tasks doesn’t justify the 3x price premium anymore.