Anthropic shipped Opus 5 at the same $5/$25 it charged for 4.8. The story is what that price now buys: half the Fable 5 bill at close to Fable 5 scores.
Claude Opus 5 landed on July 24 with a rate card copied straight from Opus 4.8, $5 input and $25 output per million tokens. For the fourth Opus release running, the list price did not move. What moved is everything behind it. SWE-bench Verified jumps from 88.6 to about 96, SWE-bench Pro from 69.2 to 79.2, and Opus 5 has quietly edged past Anthropic's own Fable 5 flagship to the top of the Artificial Analysis intelligence index while costing exactly half as much to run. Here is the full price card, where the benchmarks actually moved, and the cost math that makes Opus 5 awkward for Fable 5.
Photo by Conrad Crawford on Unsplash
Two things before the tables
- Opus 5 costs what Opus 4.8 cost, $5 in and $25 out per million, same 1M context and 128K output ceiling, but Anthropic is pitching it as near-Fable-5 intelligence, and Fable 5 lists at exactly double the rate. So Opus 5 does the flagship job at half the invoice.
- Watch the thinking toggle. Reasoning is on by default and you cannot switch it off at the top two effort levels, so the highest tiers always meter reasoning tokens at the $25 output rate.
The price card, unchanged for a fourth release
There is not much drama in the rate card, and that is the point. Standard, cache, and batch all carry straight over from Opus 4.8. Fast mode keeps the $10/$50 preview number that 4.8 introduced. One small change worth noting for prompt-caching setups: the cache minimum drops to 512 tokens, so shorter shared prefixes now qualify for the 90% cache-read discount that used to need a 1,024-token floor.
| Tier | Input / 1M | Output / 1M | Notes |
|---|---|---|---|
| Standard | $5.00 | $25.00 | Same as Opus 4.8. 1M context, 128K max output. |
| Fast mode (preview) | $10.00 | $50.00 | 2.5x speed. API-only, not on Bedrock, Vertex, or Foundry. |
| Cache read | $0.50 | n/a | 90% off input on hits. New 512-token minimum. |
| Cache write (5 min) | $6.25 | n/a | 1.25x input, default cache TTL. |
| Batch API | $2.50 | $12.50 | 50% off standard for async jobs. |
Anthropic points the cache and batch lines at its general pricing page rather than restating exact cents on the Opus 5 doc, so if you are budgeting a cache-heavy pipeline to the dollar, confirm the write and read rates there. The base $5/$25 is stated outright, and it matches what LiteLLM and the cloud resellers list.
Where the numbers actually moved
The Opus 4.8 launch two months ago moved most of its benchmarks by a point or two. Opus 5 is a wider jump, concentrated in agentic coding and the frontier-reasoning boards Anthropic has been leaning on. The figures below come from the System Card and the launch post; treat the exact decimals as Anthropic-reported until Artificial Analysis and the independent harnesses re-run them.
| Benchmark | Opus 5 | Opus 4.8 | Delta |
|---|---|---|---|
| SWE-bench Verified | 96.0 | 88.6 | +7.4 |
| SWE-bench Pro | 79.2 | 69.2 | +10.0 |
| Frontier-Bench v0.1 | 43.3 | 18.7 | +24.6 |
| ARC-AGI-3 | 30.2 | 1.5 | +28.7 |
| GPQA Diamond | ~93 | 93.6 | flat, saturated |
Read the table with two caveats. GPQA Diamond is not a real regression, both scores sit inside the noise of a benchmark that is close to maxed out, so ignore the sub-point gap. And ARC-AGI-3 going from 1.5 to 30.2 looks dramatic because the previous number was essentially a floor; the honest read is that Opus 4.8 could barely touch that benchmark and Opus 5 can, not that it is now solved. Where the jump is real money is SWE-bench Pro, the closest public proxy for the messy multi-file agent work that runs up a tokencost reader's bill.
The Fable 5 problem: same job, half the invoice
Anthropic sells two frontier tiers now. Fable 5 sits at the top at $10/$50, and Opus 5 slots in underneath at $5/$25. Those two rate cards are not just close, they are an exact 2x apart on both input and output. That has a clean consequence: whatever your read-to-write ratio, Opus 5 always comes out at exactly half the Fable 5 bill. There is no workload mix where the gap narrows or widens. Below are three real shapes at list rates, no cache or batch, against Fable 5, GPT-5.6 Sol, and Kimi K3 for context.
| Workload | Opus 5 | Fable 5 | GPT-5.6 Sol | Kimi K3 |
|---|---|---|---|---|
| Coding session (500K in / 150K out) | $6.25 | $12.50 | $7.00 | $3.75 |
| Whole-repo review (1M in / 50K out) | $6.25 | $12.50 | $6.50 | $3.75 |
| Quick chat reply (10K in / 2K out) | $0.10 | $0.20 | $0.11 | $0.06 |
| Monthly agent spend (100M in / 30M out) | $1,250 | $2,500 | $1,400 | $750 |
The GPT-5.6 Sol column is the interesting one for anyone not locked to Anthropic. Sol matches Opus 5 on input at $5 but charges $30 on output versus $25, so on output-heavy agent work Opus 5 quietly undercuts it. Kimi K3 is far cheaper still and now ranks above Opus 4.8 on some boards, which is the argument we made in the value-frontier piece. Opus 5 does not erase that gap, but it does reset the Anthropic side of it: you are no longer paying Fable 5 money for a flagship.
One behavioral change that can cost you
Opus 5 ships with an effort ladder, low through max, and reasoning is on by default. Anthropic frames this as a cost-and-capability toggle: dial effort down for cheap, fast answers, dial it up for the hard agent runs. The catch that touches your bill is that you can only disable thinking at effort high or below. Ask for xhigh or max with thinking off and the API returns a 400. So the top of the ladder always burns reasoning tokens, which bill at the $25 output rate.
In practice that means the effort dial is your real cost control on Opus 5, more than the tier is. A max-effort run on a 1M-context task can emit a lot of invisible thinking before the visible answer, and all of it meters. If you were running Opus 4.8 with thinking off for latency, audit those calls before you swap the model ID, because the same config may not be legal at the effort level you were using.
Should you move off Opus 4.8, or off Fable 5?
Coming from Opus 4.8, this is close to a free upgrade. Same price, same 1M context, same tokenizer, same cache header surface, and a genuine coding jump. The swap is claude-opus-4-8 to claude-opus-5 and nothing else in your code needs to move, apart from the thinking-toggle check above. If your Opus 4.8 bills have been steady, expect the new bills to look the same with better output.
Coming from Fable 5 the call is harder to make in Fable 5's favor than it was a week ago. Opus 5 has actually taken the #1 spot on the Artificial Analysis intelligence index, 61 to Fable 5's current 60, so the flagship you are paying double for no longer clearly leads on the headline aggregate. Fable 5 still has an argument in its highest-effort launch configuration and on a handful of the hardest frontier evals, but for the broad middle of agentic coding, paying twice the rate for the last point or two is hard to justify. Run your own token mix through the calculator before you decide, since the gap is a straight 2x and it compounds fast at scale.
Sources
- Anthropic: Introducing Claude Opus 5 - Launch announcement, July 24 2026
- Anthropic: What's new in Opus 5 - Effort ladder, thinking defaults, 512-token cache minimum
- Anthropic: Claude pricing - Standard $5/$25, Fast $10/$50, cache and batch rates
- CNBC: Anthropic launches Opus 5 - Half-the-cost-of-Fable-5 positioning
- VentureBeat: Opus 5 launch coverage - Coding-agent framing and availability
- Vellum: Opus 5 benchmarks explained - SWE-bench Verified 96.0, SWE-bench Pro 79.2, System Card readout