Four models, four price points, and a roughly 10x spread from the cheapest to the most expensive per token: Haiku 4.5 at $1 input / $5 output per million tokens, Sonnet 5 at $3/$15 (discounted to $2/$10 through August 31, 2026), Opus 5 at $5/$25, and Fable 5 at $10/$50. That headline list, though, is only the starting point — prompt caching, batch processing, and Opus 5's fast mode all move the real cost of running these models by a meaningful margin, in directions the sticker price doesn't show.
What changed in 2026
- The whole lineup refreshed to the 5 generation, except Haiku. Fable 5, Opus 5, and Sonnet 5 all launched this year; Haiku 4.5 remains the value tier without a numbered successor yet.
- Sonnet 5 shipped with a genuine introductory discount, not just a marketing framing — $2 input / $10 output per million tokens through August 31, 2026, versus $3/$15 afterward.
- Opus 5 held its predecessor's price exactly, making it one of the rare "same cost, more capability" upgrades in the lineup rather than a price increase.
- Fable 5 launched above Opus 5 on price, reflecting its narrower, harder use case rather than being positioned as a blanket upgrade.
- Fast mode introduced a new premium tier on top of Opus 5, at roughly double the standard Opus 5 rate, for teams that need faster output more than they need a cheaper one.
Full pricing table
| Model |
Input (per million tokens) |
Output (per million tokens) |
Context window |
Max output |
| Claude Haiku 4.5 |
$1.00 |
$5.00 |
200K |
64K |
| Claude Sonnet 5 |
$3.00 ($2.00 intro through Aug 31, 2026) |
$15.00 ($10.00 intro) |
1M |
128K |
| Claude Opus 5 |
$5.00 |
$25.00 |
1M |
128K |
| Claude Opus 5 (fast mode) |
$10.00 |
$50.00 |
1M |
128K |
| Claude Fable 5 |
$10.00 |
$50.00 |
1M |
128K |
Fast mode and Fable 5 land at the same headline rate, which is a coincidence worth noting rather than a signal that they're interchangeable — fast mode is a speed upgrade on Opus 5's existing capability, while Fable 5 is a different, more capable model entirely.
The levers that actually change your bill
- Prompt caching. Reusing a large system prompt, document, or tool definition set across requests means only the first call pays full price; cached reads run at roughly a tenth of the standard input rate. This is the single biggest lever for agentic and RAG-heavy workloads with a stable prefix.
- Batch processing. Non-latency-sensitive work — bulk classification, offline summarization, large-scale extraction — runs at half the standard per-token price when submitted as a batch instead of individual real-time requests.
- Model routing. Sending only the requests that need it to Opus 5 or Fable 5, and defaulting everything else to Sonnet 5 or Haiku 4.5, typically has a bigger effect on total spend than any single per-token discount.
- Effort level. Lower effort settings reduce thinking and tool-call volume, which shows up directly in output token counts on agentic tasks — this is a lever independent of which model you've chosen.
Common mistakes
- Comparing models on input price only. Output tokens dominate the bill on coding, long-form writing, and agentic tool-calling — always model both sides before picking a tier on price.
- Missing the Sonnet 5 discount window. Teams that priced out Sonnet 5 at $3/$15 and passed on it are quoting a higher number than what's actually billed through August 31, 2026.
- Ignoring prompt caching until costs are already a problem. It's one of the easiest changes to make and one of the most commonly skipped, especially in early prototypes that later scale into production traffic.
- Assuming fast mode and Fable 5 solve the same problem. Fast mode buys speed on Opus 5's existing quality; Fable 5 buys more capability at a similar price point. Picking the wrong one wastes the premium.
FAQ
Which model is cheapest for high-volume, simple tasks?
Haiku 4.5, at $1 input / $5 output per million tokens — by a wide margin the cheapest current-generation option.
Is the Sonnet 5 introductory price permanent?
No, it runs through August 31, 2026. After that date, standard pricing of $3 input / $15 output per million tokens applies.
Does prompt caching work the same way across all four models?
The mechanism is the same, but the minimum prompt length required to cache differs by model — Opus 5 and Fable 5 have a lower minimum than Sonnet 5, so shorter prompts are more likely to qualify on the top tiers.
Is Fable 5 ever cheaper than Opus 5 in practice?
Rarely on a per-token basis, since its rate is roughly double Opus 5's. It can still be cheaper overall if it solves a hard task in one pass where Opus 5 would need multiple retries.
Where to go next
For a deeper look at each tier individually, read our Claude Sonnet 5 review and Claude Opus 5 review. If you're deciding between the two mid-to-top tiers specifically, see Sonnet 5 vs Opus 5 for real workloads.