Anthropic · Pricing
Claude Opus 5 pricing
Opus 5 shipped on July 24, 2026 at $5 in / $25 out per 1M tokens — the same rate Opus 4.8 charged. Anthropic did not cut the price. It is also, by Artificial Analysis' measurement, now the highest-scoring model on their Intelligence Index. Both of those things are true at once, which is what makes the coverage confusing.
Rates and figures checked 2026-07-25 against Anthropic's launch post and Artificial Analysis' independent evaluation.
The rate card
| Mode | Input / 1M | Output / 1M |
|---|---|---|
| Opus 5 | $5.00 | $25.00 |
| Opus 5 (Fast Mode) | $10.00 | $50.00 |
| Opus 4.8 (predecessor) | $5.00 | $25.00 |
- Fast Mode is derived, not quoted. Anthropic publishes it as a multiplier — “twice Opus 5's base price on the Claude Platform” — rather than as its own rate card. The $10 / $50 above is that multiplier applied. Treat it as arithmetic, not as a quoted figure.
- No context window published. Anthropic's launch post does not state one, so this page does not carry one. If you see a figure quoted elsewhere, it has almost certainly been borrowed from Opus 4.8.
- No cached-input or batch rates published in the launch post either. Absence here means we did not find it, not that it does not exist.
Three different claims that all sound like “cheaper”
Several outlets ran Opus 5 as arriving “at half the price of Anthropic's most powerful model”. That is not wrong, but it is not about the number you are billed. Three separate cost claims are in circulation and they answer three different questions.
1. The list rate: unchanged.
$5 / $25 per 1M, identical to Opus 4.8. Anthropic's own framing is "greatly improved performance for the same cost as its predecessor." If you are already on Opus 4.8, your per-token bill does not move.
2. Anthropic's benchmark cost claims: where “half” comes from.
These are cost-to-complete comparisons against Fable 5 on specific benchmarks, published by Anthropic: within 0.5% of Fable 5's peak CursorBench 3.2 score “at half the cost”, and beating Fable 5 on OSWorld 2.0 “at just over a third of the cost”. Real claims, vendor-run, benchmark-specific — and not a statement about either model's rate card.
3. Independent cost per task: −26%.
Artificial Analysis, evaluating separately, puts Opus 5 (max) at $2.03 per Intelligence Index task against $2.75 for Fable 5 with fallback. On their AA-Briefcase agentic benchmark the gap is $17.79 versus $22.30, about 20% lower. This is the closest thing to an apples-to-apples answer, and it is the one worth quoting.
The practical translation: Opus 5 is not cheaper than the model you were using if that model was Opus 4.8 — it is better at the same rate. It is cheaper than Fable 5 for equivalent work, which is the comparison Anthropic is making and the one the headlines compressed.
Fewer tokens, bigger bill: the arithmetic
Here is the number that should be interesting and mostly is not being examined. Artificial Analysis reported what it cost them to run each model through the Intelligence Index:
| Model | Cost to evaluate | Tokens generated | Index score |
|---|---|---|---|
| Claude Opus 5 (max) | $3,835.51 | 100M | 61 |
| Claude Opus 4.8 (max) | $3,752.55 | 120M | 56 |
Same vendor, same rates, same evaluation suite. Opus 5 generated 20 million fewer tokens and still cost $82.96 more to run. Coverage has noted the oddity and left it there. It resolves with one subtraction.
Where the money actually went
- Output at $25/1M: Opus 5's 100M costs $2,500. Opus 4.8's 120M costs $3,000. On the output side Opus 5 was $500 cheaper.
- The totals differ by +$82.96 against Opus 5. So roughly $583 of additional cost has to sit somewhere other than generated tokens — and on a two-line rate card, the only other line is input. At $5/1M that is on the order of 116 million extra input tokens.
- The likely mechanism, and here we are reasoning rather than reading a figure: in agentic and multi-turn evaluation the whole conversation is re-sent on every turn, so input volume tracks turn count, not verbosity. A model that gets further into hard tasks takes more turns, and pays for the transcript again each time. Terseness in the reply does not buy you out of that.
- Where the trail ends: Artificial Analysis has not published the input/output split for either run, so the $583 is bounded by subtraction rather than confirmed by a line item. We are labelling that rather than rounding it into a conclusion.
The figure that survives all of this is the normalised one: each Index point cost about 6.2% less on Opus 5 than on Opus 4.8. That is the honest version of “more efficient” — better value per unit of measured capability, on a bill that went up rather than down. Both models were flagged as very verbose against the 63M cross-model average, so neither is a frugal choice in absolute terms.
The setting that moves your bill more than the rate does
Because the rate is fixed, the only cost lever you control is the effort setting — Anthropic describes it as letting you “optimize for intelligence or conserve tokens for faster and cheaper results.” Every headline figure on this page, including the 61 and the $2.03, is the max effort variant. Artificial Analysis tracks the lower efforts as separate entries, which means a quoted Opus 5 number is meaningless unless you know which effort it came from.
Fast Mode is the other lever and it runs the opposite way: it doubles the rate to buy latency. It is a spend decision, not a savings one, and worth separating from the effort dial when you are budgeting.
Where it sits, as of 2026-07-25
| Model | Intelligence Index | GDPval-AA v2 Elo |
|---|---|---|
| Claude Opus 5 (max) | 61 | 1861 |
| Claude Fable 5 (max) | 60 | ~1747 |
| GPT-5.6 Sol (max) | 59 | — |
| Kimi K3 | 57 | — |
| Claude Opus 4.8 (max) | 56 | — |
Opus 5 (max) also leads AA-Briefcase at 1720 Elo, +146 over Fable 5, and matches Fable 5 at 53% on Humanity's Last Exam. Fable 5's GDPval figure is derived from the published +114 gap rather than quoted directly, hence the tilde.
This launch is why another page on this site changed
Opus 5's arrival — with its effort variants counted as separate entries — is part of what pushed Kimi K3 from #4 to #7 on the Intelligence Index in nine days, without K3's score changing by a point. It is a clean illustration of why we quote scores rather than ranks: a rank is a position in a field that other labs keep moving.
What your own volume costs
Rate-card arithmetic on the base Opus 5 tier against the frontier models we track. This is list price on your token counts — it does not model the re-sent-context effect described above, which is exactly the gap between a rate card and a bill.
Cost calculator
Enter your monthly token volume to see what each tier would cost.
| Tier | Input cost | Output cost | Monthly total |
|---|---|---|---|
| Opus 5 | $5.00 | $5.00 | $10.00 |
| Sol | $5.00 | $6.00 | $11.00 |
| Kimi K3 | $3.00 | $3.00 | $6.00cheapest |
Estimates only — list prices, before any caching discount, batch pricing, or enterprise agreement.
Sources
- Anthropic — Introducing Claude Opus 5 (launch post, rates and Fast Mode) — retrieved 2026-07-25
- Artificial Analysis — Opus 5: Fable 5 level intelligence at a lower cost per task — retrieved 2026-07-25
- Artificial Analysis — Claude Opus 5 (max) model page — retrieved 2026-07-25
- Implicator.ai — Opus 5 cut tokens 17% but cost more to benchmark than Opus 4.8 — retrieved 2026-07-25
Figures on this page last checked against these sources on 2026-07-25. Vendors change pricing and specs without notice — if a number here disagrees with the vendor's own page, trust the vendor.