Comparison
Claude Opus 5.5 vs Claude Opus 5: Should You Upgrade?
EvoLink Team
Product Team
September 22, 2026
15 min read
Short answer: pilot Claude Opus 5.5 for long-running agentic coding, large-document analysis, and chart-heavy visual work, and keep Claude Opus 5 as the rollback route until a matched replay says otherwise. Opus 5.5 costs 20% less per token on Anthropic's list price, 60% less on cache reads, and Anthropic says it finishes tasks with fewer tokens. It is not a drop-in swap: four request patterns that work on Opus 5 return a
400 on Opus 5.5, the default effort drops from high to medium, and the model thinks more per turn at a given effort level.This guide is for teams that already run Opus 5 in production. Release facts and how the "Opus 5.2" rumor resolved are in Claude Opus 5.5 release date. Live EvoLink rates, the model ID, and integration settings are on the Claude Opus 5.5 model page and the Claude Opus 5 model page. Prices below are Anthropic's published list prices, used only to compare the two models on the same basis.
The Decision at a Glance
| If your situation is | Do this | Because |
|---|---|---|
| New workload, no Opus baseline yet | Start on Opus 5.5 | Anthropic's recommended default, lower list price, same limits |
Opus 5 in production, thinking already on, tool_choice: auto | Pilot 5.5 behind a canary | Code usually carries over once effort is set explicitly; still check the five risks below (thinking settings, tool choice, progress text, thinking-block fallback, computer-use version) |
Opus 5 in production with thinking: disabled or forced tool choice | Fix the requests first, then pilot | Both patterns return 400 on 5.5 |
| Cache-heavy agent with a large stable prefix | Prioritize the pilot | Cache reads are 60% cheaper on 5.5 |
| Fallback chain that drops from Opus to Sonnet or Haiku mid-conversation | Re-test the fallback path | Only Fable 5.1 and Mythos 5.1 read Opus 5.5 thinking blocks |
| Streaming UI that shows text between tool calls | Add thinking.display handling before rollout | That text now arrives in thinking blocks, empty by default |
| Latency-sensitive, thinking was off for speed | Compare low effort on 5.5 against Opus 5 with thinking off | Thinking cannot be disabled; effort is the only lever |
There is no unconditional winner in this table. The upgrade pays when your workload is one Anthropic optimized for and your request shape is already compatible; it costs engineering time first when it is not.
What Anthropic Changed
| Dimension | Claude Opus 5 | Claude Opus 5.5 | Documented by |
|---|---|---|---|
| Release | July 24, 2026 | September 22, 2026 | Release notes |
| Model ID | claude-opus-5 | claude-opus-5-5 | Models overview |
| Input / output | $5 / $25 per MTok | $4 / $20 per MTok | Pricing |
| Cache write 5m / 1h | $6.25 / $10 | $5 / $8 | Pricing |
| Cache read | $0.50 (10% of input) | $0.20 (5% of input) | Pricing |
| Batch | $2.50 / $12.50 | $2 / $10 | Pricing |
| Fast mode (Claude API only) | $10 / $50 | $8 / $40 | Pricing |
| Context / max output | 1M / 128K | 1M / 128K | Models overview |
| Thinking | On by default; disabled allowed at effort high or below | Always on; disabled and budget_tokens return 400 | What's new |
| Default effort | high | medium | Models overview |
Forced tool choice (any, tool) | Supported | Returns 400 | What's new |
| Text between tool calls | text blocks | Progress-update thinking blocks, empty at default display | What's new |
| Thinking-block portability | Read by Opus 5.5, Fable 5.1, Mythos 5.1 | Read only by Fable 5.1 and Mythos 5.1 | What's new |
| Computer use on Claude API / Google Cloud | Toolset or computer_20251124 | Toolset only | What's new |
| Safeguard categories | cyber | cyber, bio, reasoning_extraction | What's new |
| Knowledge cutoff | Earlier | June 2026 | Models overview |
| Retirement floor | Not before July 24, 2027 | Not before September 22, 2027 | Models overview |
Everything in this table is a documented contract change you can verify. Anthropic's capability statements are a separate category: it reports leading FrontierCode results, parity with Claude Fable 5.1 "on most work," more than 30% faster output, and 40% lower running cost than Opus 5. Those are claims about Anthropic's task mix. Treat them as the hypotheses your evaluation tests.
Cost on the Same Basis: 20%, 60%, or 40%?
Three different numbers circulate, and they measure different things.
- 20% is the list-price gap on input and output. It applies to every token, before any behavior difference.
- 60% is the list-price gap on cache reads. It applies only to the share of your input that is served from cache.
- 40% is Anthropic's estimate of total running cost on its own workloads. It bundles the unit-rate cuts with fewer tokens per completed task, so it depends on how your tasks behave on the new model.
The table below applies the list prices to three fixed token mixes, per 1,000 tasks. It is a price calculation, not a quality-adjusted cost.
| Workload (per task) | Opus 5 | Opus 5.5 | Change |
|---|---|---|---|
| A. Single-shot, no cache: 20K input, 4K output | $200.00 | $160.00 | -20% |
| B. Cache-heavy agent: 60K cache reads, 10K fresh input, 5K cache write, 6K output | $261.25 | $197.00 | -25% |
| C. Same as A, but 5.5 emits 25% more output tokens at the same effort | $200.00 | $180.00 | -10% |
Workload C is the one to watch. Anthropic's documentation says Opus 5.5 tends to think more per turn than Opus 5 at the same effort level, most of all at
xhigh and max, and thinking is billed as output. On mix A, Opus 5.5 can emit up to 50% more output tokens before its per-task price matches Opus 5. Whether it does depends on the effort level you choose and on whether the model reaches an accepted result in fewer turns, which is where Anthropic's 40% figure comes from.The only comparison that settles an upgrade is cost per accepted task: total spend on the evaluation group, including failures, retries, and tool fees, divided by the number of tasks that passed acceptance, with human fix time reported separately. Run it after the effort sweep in the next section, not before.
Compatibility Risks, in the Order They Bite
1. Requests that now fail
Scope note: the
400 errors below are Anthropic API behavior. On EvoLink, the gateway converts legacy thinking: disabled and budget_tokens settings to adaptive thinking instead of rejecting them, while forced tool choice is rejected on both the Chat Completions and Messages endpoints; the Claude Opus 5.5 model page documents the gateway behavior.Search your request builders for three patterns.
thinking: {"type": "disabled"} and thinking: {"type": "enabled", "budget_tokens": N} both return 400 invalid_request_error. tool_choice: {"type": "any"} and {"type": "tool", "name": ...} return 400, including on the token-counting endpoint. On the Claude API and Google Cloud, a computer_20251124 tool entry returns 400; declare computer_toolset_20260801 instead.The replacements are mechanical. Remove the
thinking field, or send thinking: {"type": "adaptive"}, which is equivalent, and set output_config.effort explicitly. Keep tool_choice: auto, set strict: true on tools that must produce schema-valid arguments, or move the schema to structured outputs, and say in the prompt when the tool applies. Then verify in your loop that a tool call actually occurred, because auto does not guarantee one.2. The default that moves silently
A request that omits
effort ran at high on Opus 5 and runs at medium on Opus 5.5. If your production requests never set effort, your first Opus 5.5 replay is not a like-for-like comparison. Set the level explicitly on both models before you read any result, then sweep low, medium, and high on 5.5.3. Progress text that disappears
Between tool calls, Opus 5 returned short narration as
text blocks. Opus 5.5 returns it as progress-update thinking blocks, at most one before each tool call, with an empty thinking field at the default display: "omitted". Nothing errors; your streaming UI just goes quiet. Set thinking.display to "updates" (beta header thinking-display-updates-2026-08-18) or "summarized", render each non-empty thinking block ahead of the tool call it precedes, and pass the blocks back unchanged.4. Fallback chains that lose their reasoning
Thinking blocks record the model that produced them. Opus 5.5 reads blocks from Opus 5 and earlier Opus, Sonnet, and Haiku models, so a conversation that starts on Opus 5 and moves to 5.5 keeps its reasoning. In the other direction, only Claude Fable 5.1 and Mythos 5.1 read Opus 5.5 blocks. A router that falls back from Opus 5.5 to Opus 5, Sonnet 5, or Haiku 4.5 runs the remaining turns without the earlier reasoning; the request succeeds and the dropped blocks are not billed, but the behavior changes. The API also checks that nothing before a 5.5 thinking block was edited; accounts created on or after August 31, 2026 get a
400 by default when they replay a block after such an edit. Keep history append-only and use mid-conversation system messages instead of editing the system prompt or tools.5. New refusal categories
Opus 5.5 runs a biology classifier alongside the cybersecurity one and can decline requests that push it to reproduce internal reasoning, returning
stop_reason: "refusal" with stop_details.category set to bio or reasoning_extraction. Server-side fallback does not retry reasoning_extraction. If your Opus 5 integration never inspected stop_reason, add that check before the pilot.Six Scenarios and What Each Must Prove
| Scenario | Input to output | What "better" means | What to watch on Opus 5.5 |
|---|---|---|---|
| Multi-file code change | Repository context to patch | Tests pass; reviewer accepts with fewer edits | Output tokens at the chosen effort; steps to green |
| Multi-turn tool agent | Task plus tool schemas to a sequence of tool calls | Task completed; loop recoverable after failure | auto still calls the tool; progress blocks rendered; thinking blocks passed back |
| Long-document analysis | Documents and images to sourced conclusions | Claims match the evidence set; no missed items | Citation checks; Anthropic reports fewer unsupported figures, verify it |
| Structured extraction | Semi-structured text to a fixed JSON shape | Schema-valid rate; field accuracy | strict: true replaces forced tool choice |
| Cache-heavy workflow | Fixed prefix plus a new task to a result | Acceptance rate holds; the bill actually falls | Cache hit share; 5m versus 1h write choice |
| Multi-model fallback | Failed or timed-out turn to an alternative route | Conversation continues; cost stays bounded | Which models can read 5.5 thinking blocks; refusal categories |
None of these are results. They are the acceptance records to fill in before traffic moves, one per task class, with the same prompts, tools, timeouts, effort levels, retry rules, and reviewers on both models.
When to Keep Opus 5
Keep Opus 5 where it already meets your pass-rate, latency, and cost targets and the replay shows no measured upside. Keep it where your request builders still send
thinking: disabled or forced tool choice and the fix has not shipped. Keep it as the rollback route for every 5.5 canary until the canary has run long enough to see p95 latency and refusal behavior under real load. And keep it where a fallback chain depends on downstream models reading the conversation's thinking blocks, until that chain is re-tested.Waiting is not passive while you are collecting baselines. Anthropic lists Opus 5 as active with a retirement floor of July 24, 2027 on the platforms it operates, so there is time to move on evidence rather than on the release date.
A Safe Migration Plan
1. Freeze the Opus 5 baseline
Pick representative tasks from each class above. Record the exact request shape, effort level, tool set, and acceptance criteria, and run them on Opus 5 once more so the baseline is fresh and not a memory.
2. Fix the request shape
Remove
thinking: disabled and budget_tokens, set effort explicitly, replace forced tool choice with auto plus strict, move computer use to the toolset, and add stop_reason handling. Ship this to Opus 5 first; every change is valid there too, which lets you separate the request fix from the model change.3. Run the effort sweep
Replay the baseline on Opus 5.5 at
low, medium, and high. Record accepted-task rate, output tokens, wall time, and cost per accepted task at each level. Pick the cheapest level that meets the acceptance bar, per task class. Do not carry high over from Opus 5 by habit.4. Canary by task class
Route a small share of one task class to Opus 5.5 with Opus 5 as the configured fallback, and log requested model, returned model, effort,
stop_reason, and billed dimensions on every call. Promote a class when its cost per accepted task and quality both clear the gate for a full cycle of real traffic, not a single good day.5. Define the way back
Rollback is a configuration change: switch the class back to
claude-opus-5. Because Opus 5 cannot read Opus 5.5 thinking blocks, a mid-conversation rollback continues without the earlier reasoning; start a fresh conversation boundary when you can. Record who can trigger the rollback and what signal triggers it before the canary starts.On EvoLink the model selection is a string in your request, so steps 4 and 5 do not require a code deploy.
Review the Claude Opus 5.5 route and current rates Keep the Claude Opus 5 route as your rollbackFAQ
Is Claude Opus 5.5 cheaper than Claude Opus 5?
Per token, yes: Anthropic lists input and output 20% lower and cache reads 60% lower. Per task, it depends on how many tokens the model spends at your effort level; Anthropic's 40% figure is its own estimate on its own workloads.
Is Opus 5.5 a drop-in replacement for Opus 5?
No. Requests that disable thinking, set a thinking budget, force a tool choice, or declare the older computer-use tool return a
400. The default effort also drops from high to medium, and text between tool calls moves into thinking blocks.Why might Opus 5.5 cost more than Opus 5 on my workload?
Because it tends to think more per turn at the same effort level, and thinking is billed as output. On a 20K-input, 4K-output task it can emit 50% more output before its price matches Opus 5, so choose effort from a sweep rather than carrying
high over.Can I turn thinking off on Opus 5.5 for latency?
No. Use
effort: low and compare it against your Opus 5 thinking-off baseline on latency and acceptance. Anthropic's guidance for prompts written for thinking-off also suggests removing "do not reason" instructions.What happens to thinking blocks if I fall back from Opus 5.5 to Opus 5?
Opus 5 cannot read them, so the API drops them before the model sees the request. The request succeeds and the dropped blocks are not billed, but the remaining turns run without the earlier reasoning. Going from Opus 5 to Opus 5.5 keeps them.
Does forced tool choice have a replacement?
Keep
tool_choice: auto, set strict: true on the tool so any call matches its schema, and state in the prompt when the tool applies. For pure JSON output, use structured outputs instead of a tool.Should Opus 5 stay available as a fallback?
Yes, at least through the canary. Anthropic lists
claude-opus-5 as active with a retirement floor of July 24, 2027 on the platforms it operates. Test the specific route you plan to fall back to, including the thinking-block behavior above.Which workloads should move first?
Long agentic coding runs, large-document analysis, and cache-heavy agents with a stable prefix, in that order of expected payoff. Everyday coding and high-volume, latency-sensitive traffic usually belong on Sonnet 5 rather than on either Opus.
Sources
- Anthropic: What's new in Claude Opus 5.5
- Anthropic: Migrating to Claude Opus 5.5
- Anthropic: Models overview
- Anthropic: Pricing
- Anthropic: Effort parameter
- Anthropic: Preserved thinking
- Anthropic: Refusals and fallback
- Anthropic: Introducing Claude Opus 5.5
Facts last verified September 22, 2026. Cost examples use Anthropic list prices and fixed token mixes; they are price arithmetic, not measured results.
