Anthropic released Claude Opus 5 on 24 July 2026 at US$5 per million input tokens and US$25 per million output — the same rate card as Opus 4.8, and exactly half the price of its top-end Fable 5.
That framing matters. This is not a price cut. It is a capability claim at a held price: more work per dollar, plus an effort control that trades depth for cost on each request. So we did what the launch coverage didn’t — priced it against ten rivals from the providers’ own rate cards.
What Anthropic claims — all of it company-reported
The launch material sells cost-per-task, not raw scores. Five claims carry the pitch:
| Benchmark | The claim (Anthropic’s own numbers) |
|---|---|
| CursorBench 3.2 | Within 0.5% of Fable 5’s peak at max effort — at half the cost per task |
| OSWorld 2.0 | Above Fable 5’s best result at just over a third of the cost |
| Frontier-Bench v0.1 | More than double Opus 4.8’s result, at lower cost per task |
| ARC-AGI 3 | Three times the next-best model’s score |
| Zapier AutomationBench | Around 1.5× the next-best pass rate at the same cost per task |
Every figure above is Anthropic measuring Anthropic, on Anthropic’s harness. None had been independently replicated when we published, about a day after launch — and the per-task arithmetic depends on how many tokens the model burns at each effort setting, a dial the vendor controlled in every cited run.
Five releases, one price
The rate card is the quiet flex. Anthropic’s own pricing table tells the story:
-
The US$15 / US$75 era
Opus tops the price list; the tier is later deprecated at the same rate.
-
The big drop: US$5 / US$25
Opus pricing falls by two-thirds and the modern rate card is born.
-
Held. Held. Held.
Three consecutive releases ship capability gains at the identical price.
-
Opus 5 holds it again
A fifth release at US$5 / US$25 — now positioned as half of Fable 5.
Fast mode — research preview, first-party API only — doubles the price for faster output. Batch processing halves it. Cache reads run at a tenth of input. The discipline runs through the whole card.
We ran the numbers
All prices below come from the five providers’ official pricing pages, fetched 25 July 2026. Blended cost per million tokens is (r × input + output) ÷ (r + 1): 3:1 is chat-shaped, 10:1 is agent-shaped — context, tool schemas and retries make input dominate. Rerun any row against your own workload with the LLM Price Comparison, or project a monthly bill with the Monthly LLM Spend Projector.
| Model | In / Out $/M | Blend 3:1 | Blend 10:1 | Cached in |
|---|---|---|---|---|
| Opus 5 | $5.00 / $25.00 | $10.00 | $6.82 | $0.50 |
| Fable 5 | $10.00 / $50.00 | $20.00 | $13.64 | $1.00 |
| Sonnet 5 ¹ | $2.00 / $10.00 | $4.00 | $2.73 | $0.20 |
| GPT-5.6 Sol | $5.00 / $30.00 | $11.25 | $7.27 | $0.50 |
| GPT-5.6 Terra | $2.50 / $15.00 | $5.63 | $3.64 | $0.25 |
| GPT-5.6 Luna | $1.00 / $6.00 | $2.25 | $1.45 | $0.10 |
| Gemini 3.1 Pro ² | $2.00 / $12.00 | $4.50 | $2.91 | $0.20 |
| Gemini 3.6 Flash | $1.50 / $7.50 | $3.00 | $2.05 | $0.15 |
| Grok 4.5 ² | $2.00 / $6.00 | $3.00 | $2.36 | — |
| DeepSeek V4-Pro | $0.435 / $0.87 | $0.54 | $0.47 | $0.0036 |
| DeepSeek V4-Flash | $0.14 / $0.28 | $0.18 | $0.15 | $0.0028 |
Blended $/M = (r × input + output) ÷ (r + 1). Prices read from each provider's official pricing page on 25 July 2026.
¹ Introductory pricing, announced at launch as running to 31 August 2026. Correction, 14 August 2026: Anthropic made this rate permanent on 10 August; the scheduled rise to US$3 / US$15 on 1 September will not occur. ² Rates roughly double on prompts above 200k tokens.
The half-price claim checks out per token — everywhere. Opus 5 is exactly 50% of Fable 5 on base input, base output, cache reads and batch rates. Whether it is half the cost per task depends on token burn at each effort setting — the part only Anthropic has measured.
Against OpenAI, the gap is output. Opus 5 and GPT-5.6 Sol charge identical input; output is US$25 against US$30. At a 10:1 agent blend they sit US$6.82 to US$7.27 — close enough that token efficiency, not the rate card, decides real bills between them.
The floor is an order of magnitude down. DeepSeek V4-Pro blends to US$0.47 at 10:1 — a fourteenth of Opus 5 — and V4-Flash’s cached input, at US$0.0028 per million, is 179 times cheaper than Opus 5’s US$0.50 cache reads. Capability is a separate question — our directory tracks Claude, ChatGPT, Gemini, Grok and DeepSeek entry by entry.
The footnote that bends every comparison
Anthropic’s own pricing documentation notes that models from Claude 4.7 onward use a tokenizer producing roughly 30% more tokens for the same text than the previous one. A per-token price held flat across that change is a quiet per-text increase against older Claude models — and one more reason naive dollars-per-million comparisons across vendors, which already tokenise differently, mean less than they appear to. Paste your own text into the Tokenization Visualizer to see the spread.
The caveats that matter
- Intro pricing. Sonnet 5’s table row was announced as expiring 31 August, with US$3 / US$15 from 1 September. (Correction, 14 August 2026: Anthropic made this pricing permanent on 10 August. The rise to US$3 / US$15 on 1 September will not occur.)
- Long context. Gemini 3.1 Pro and Grok 4.5 roughly double their rates above 200k-token prompts.
- Safety positioning. Anthropic notes Opus 5 trails its limited-availability Mythos 5 line on certain security-research evaluations — a deliberate choice under its safety framing.
Key takeaways
- The price. Claude Opus 5 launched 24 July 2026 at US$5/US$25 per 1M tokens — unchanged from Opus 4.8, exactly half of Fable 5 across base, cache and batch.
- The claims. Every cost-per-task figure is company-reported and unverified at publication.
- The field. At a 10:1 agent blend, Opus 5 (US$6.82/M) edges GPT-5.6 Sol (US$7.27/M); Gemini 3.1 Pro, Grok 4.5 and both DeepSeek V4 models undercut them all.
- The deadline. Sonnet 5’s intro pricing was set to die 31 August. (Correction, 14 August 2026: it did not — Anthropic made US$2 / US$10 permanent on 10 August. Any cost model built on the higher September rate is now overstated.)
- The tokenizer. Claude’s 4.7+ models produce ~30% more tokens for the same text, reshaping every $/MTok comparison they touch.
View author profile → · Editorial policy
