Claude Opus 5 vs Claude Opus 5.5: cheaper per token, different defaults
Same vendor, same Opus line, same TulexAI plans. Anthropic’s docs now list Opus 5 as a legacy model that is still available, and it stays on TulexAI at 12 PT per 750 tokens. What changes with Opus 5.5 is the price per token, the effort each model runs at when none is set, and a stricter set of safeguards and API rules.
The usual choice
Pick one. Pay one subscription. Live inside one vendor's models.
- Claude Opus 5 at 12 PT per 750 tokens, still available
- Claude Opus 5.5 at 10 PT per 750 tokens
- Plus 54+ models, one login
Claude Opus 5.5 is Anthropic’s current Opus, listed at $4/$20 per 1M tokens against Opus 5’s $5/$25. Artificial Analysis scores it 58 to 51 at max effort on Intelligence Index v4.3.2, and 51 at medium, TulexAI’s setting. It rejects forced tool choice, and its safeguards step in more often. On TulexAI: 10 PT per 750 tokens against 12, both from Basic.
Claude Opus 5 vs Claude Opus 5.5 vs TulexAI
| Claude Opus 5 | Claude Opus 5.5 | TulexAI | |
|---|---|---|---|
| API list price per 1M tokens (input / output, cache reads) | $5 / $25, cache reads $0.50 | $4 / $20, cache reads $0.20 | 12 PT vs 10 PT per 750 tokens of prompt plus reply |
| Intelligence Index v4.3.2 at max effort (Artificial Analysis) | 51, #11 of 212 | 58, #1 of 212 | Neither runs at max effort on TulexAI |
| Intelligence Index v4.3.2 at medium effort (Artificial Analysis) | 45, #26 of 212 | 51, #8 of 212 | Medium is Opus 5.5’s setting on TulexAI; Opus 5 runs at high |
| Anthropic-reported coding scores (FrontierCode v1.1 / CursorBench 4.0) | 48.0% / 46.6% at max effort on CursorBench | 54.6% / 52.5% at medium effort | Opus 5.5’s figures are Anthropic’s, at TulexAI’s setting |
| Hallucination rate at max effort (Artificial Analysis, lower is better) | 61% | 59% | Opus 5.5 measured 68% at medium |
| Output speed and time to first token at medium effort (Artificial Analysis) | 56.6 tok/s, 4.84s | 75.2 tok/s, 22.17s | Not measured on TulexAI |
| Default effort and thinking | High | Medium, with adaptive thinking always on | No effort setting sent; thinking shares the 8,192-token reply cap |
| Tool use and forced tool choice | Tools and forced tool choice | Tools; forced choice rejected by Anthropic | Tools on both; a forced choice runs as auto on Opus 5.5 |
| Safeguard refusals in Coding Agent Index v1.5 (Artificial Analysis, max effort) | 0.5% of attempts | 9.1% of attempts | No fallback model: flagged requests are declined |
| Status at Anthropic | Legacy, still available; retirement not before July 24, 2027 | Current Opus; retirement not before Sept 22, 2027 | Both on Basic ($11/mo) and up |
Which should you pick?
Switch to Claude Opus 5.5 for most work. Per token it lists 20% below Opus 5, Artificial Analysis scores it seven points higher on Intelligence Index v4.3.2 at max effort (58 to 51), and at medium, the effort TulexAI runs it at, it scored 51, the same as Opus 5 reached at max. On TulexAI it is also the cheaper of the two: 10 PT per 750 tokens of prompt plus reply against 12. Staying on Opus 5 is reasonable in three cases: your integration forces a tool call, which Anthropic no longer allows on Opus 5.5; your prompts touch security research or biology, where Opus 5.5’s safeguards step in far more often and TulexAI declines what they flag instead of re-routing it; or you need the first token quickly: at medium effort Artificial Analysis measured 22 seconds for Opus 5.5 against under 5 for Opus 5, though on TulexAI Opus 5 runs at high, where no comparable figure has been published. Both run on the Basic plan and up, so you can send the same prompt to each before you move a workflow.
Why choose at all? TulexAI includes both
54+ models in one subscription - here are the ones that cover what Claude Opus 5 and Claude Opus 5.5 each do best.
Text & Chat
- Claude Opus 5.5Premium
- Claude Opus 5Premium
- Claude Fable 5.1Premium
Frequently asked questions
On nearly every published measure, yes. Artificial Analysis scores it 58 against 51 on Intelligence Index v4.3.2 at max effort (#1 and #11 of 212), and Anthropic’s own table puts it ahead of Opus 5 on every benchmark it lists, for example 1846 against 1708 Elo on GDPval-AA v2.1 at max effort. The exceptions are small or run the other way: its hallucination rate at max effort is only a little lower, 59% against 61%, and at medium effort it takes longer to reach its first token.
Per token, 20% less: Anthropic lists Opus 5.5 at $4 input and $20 output per 1M tokens against $5 and $25, with cache reads at $0.20 against $0.50 and 5-minute cache writes at $5 against $6.25. Anthropic also says that "at default settings it will cost 40% less than Opus 5 on typical workloads", which compares Opus 5.5 at medium effort with Opus 5 at high; that is a claim about spend per task, not a price cut. On TulexAI the difference is 10 PT against 12 per 750 tokens of prompt plus reply.
Anthropic says it generates output more than 30% faster than Opus 5, which is a claim about tokens per second. Artificial Analysis partly supports it: at medium effort it measured 75.2 tokens a second against 56.6, about 33% more. The first token arrives later, though, 22.17 seconds against 4.84, because adaptive thinking runs before the answer, and end-to-end response time was 28.83 seconds against 13.68. Those runs used Anthropic’s own API, not TulexAI.
It ships with stricter safeguards. Anthropic says bug finding and fixing in the normal development cycle stays allowed, but "most cybersecurity tasks will be re-routed to Opus 4.8". On the API a flagged request comes back as a normal response with stop_reason "refusal", and re-routing only happens when the caller opts in, with the response naming the model that served it. TulexAI does not opt in, so a flagged cyber or biology prompt is declined. In Coding Agent Index v1.5 at max effort, Artificial Analysis logged safeguard refusals on 9.1% of Opus 5.5’s attempts against 0.5% for Opus 5, close to Fable 5.1’s 8.8%.
Not soon. Anthropic’s docs call it a legacy model that is still available, with retirement not sooner than July 24, 2027, while Opus 5.5 carries a not-sooner-than date of September 22, 2027. Opus 5 remains on TulexAI at 12 PT per 750 tokens from the Basic plan, next to Opus 5.5 at 10 PT (on Basic the two share one small premium text allowance before requests downgrade), so a workflow that depends on Opus 5, for example one that forces a tool call, needs no change.
Get both - and 35+ more models
40+ AI models - GPT-5, Claude Opus 5, Gemini 3.1, image, video & voice. One subscription from $11/mo.
Try 1 prompt free1 free prompt · no credit card · cancel any time