Claude Sonnet 5: Pricing, Benchmarks, and Cost
Claude Sonnet 5 ships at $3/$15 per million tokens, intro $2/$10 through August 2026. What changed versus Sonnet 4.6 and Opus 4.8, and the cost per task.
By Capital & Compute
Anthropic released Claude Sonnet 5 on June 30, 2026, and made it the default model on the Free and Pro plans the same day. For anyone budgeting an AI workload, the headline is what did not change: the standard rate is still $3 per million input tokens and $15 per million output tokens, exactly where Sonnet 4.6 sat. The interesting part is a time-boxed discount and a capability jump that moves Sonnet into territory that used to require the Opus tier.
Here is what Sonnet 5 costs, what changed against Sonnet 4.6 and Opus 4.8, and whether any of it changes the cost per task.
How much does Claude Sonnet 5 cost?
Standard API pricing is $3 per million input tokens and $15 per million output tokens, per Anthropic’s models overview. That is identical to Sonnet 4.6. Through August 31, 2026, Anthropic is running introductory pricing of $2 / $10, stated both in the Claude Sonnet 5 launch post and as a footnote on the models page. After that date it reverts to $3 / $15.
| Model | Input / Mtok | Output / Mtok | Position |
|---|---|---|---|
| Claude Sonnet 5 (intro) | $2.00 | $10.00 | promo, through Aug 31 2026 |
| Claude Sonnet 5 (standard) | $3.00 | $15.00 | current mid-tier, from Sep 1 |
| Claude Sonnet 4.6 (legacy) | $3.00 | $15.00 | superseded, still available |
| Claude Opus 4.8 | $5.00 | $25.00 | top tier |
| Claude Haiku 4.5 | $1.00 | $5.00 | cheapest tier |
The prompt cache read stays at roughly 10% of input, about $0.30 per million tokens at the standard rate, which matters because coding agents lean heavily on cached context. For the dated, sourced rate card across every current model, the AI model tracker keeps them current.
What changed from Sonnet 4.6
The specs are a clean upgrade at the same price. Sonnet 5 carries a 1M-token context window, 128k maximum output, adaptive thinking, and a January 2026 knowledge cutoff, per Anthropic’s models overview.
| Spec | Claude Sonnet 5 |
|---|---|
| Context window | 1M tokens |
| Max output | 128k tokens |
| Adaptive thinking | Yes |
| Knowledge cutoff | January 2026 |
| API id | claude-sonnet-5 |
| Released | June 30, 2026 |
On the benchmarks Anthropic published with the launch, Sonnet 5 is a straight improvement over Sonnet 4.6 across reasoning, coding, and computer use. Anthropic presents the head-to-head as a chart in the launch post; the numbers below are that table as transcribed by DataCamp, with the two readable primary anchors (OSWorld-Verified and Humanity’s Last Exam for Sonnet 4.6) matching the launch post directly.
| Item | Sonnet 5 | Sonnet 4.6 |
|---|---|---|
| SWE-bench Pro (coding) | 63.2% | 58.1% |
| Terminal-Bench 2.1 | 80.4% | 67% |
| OSWorld-Verified (computer use) | 81.2% | 78.5% |
| Humanity's Last Exam (with tools) | 57.4% | 46.8% |
Sonnet 5 versus Opus 4.8: the value question
The reason this release matters for a budget is the gap it closes to Opus. On SWE-bench Pro, Anthropic’s harder agentic-coding benchmark, Sonnet 5 scores 63.2% against Opus 4.8 at 69.2%, per the launch benchmarks. Sonnet 5 buys most of the way to Opus-grade coding at $15 output versus Opus at $25, a 40% saving on the expensive side of the meter.
| Item | SWE-bench Pro score (%) | Output $ / Mtok |
|---|---|---|
| Sonnet 5 | 63.2% | $15 |
| Sonnet 4.6 | 58.1% | $15 |
| Opus 4.8 | 69.2% | $25 |
This is the standing trade in the Claude line: Opus for the hardest, highest-autonomy work where a few extra points of reliability pay for themselves, Sonnet for the daily-driver volume where cost per finished task decides the bill. Sonnet 5 widens the set of work where the cheaper tier is the correct call. To cost a specific matchup, the model comparison tool prices Sonnet 5 against Opus 4.8, DeepSeek V4, and the rest on the same task.
Does it change the cost per task?
At the standard rate, no. Because $3 / $15 is unchanged from Sonnet 4.6, the modeled multi-file coding task that runs about $1.46 on Sonnet costs the same on Sonnet 5. The cost-per-task math carries over unchanged, and the cost-per-task calculator now defaults to Sonnet 5 at the same figure.
Two things do move the real bill. During the intro window, that same task runs roughly a third cheaper on the $2 / $10 promo rate, so a high-volume August is the time to lean on Sonnet 5. And token efficiency matters more than the sticker: a stronger model that finishes in fewer output tokens, or with fewer retries, costs less per task even at the same rate. That is the lens worth budgeting on, the same way it applied to the GPT-5.6 pricing question.
Should you switch?
For most Claude workloads the switch already happened: Sonnet 5 is the default, and at the same standard price with better benchmarks there is no reason to pin Sonnet 4.6 unless a specific evaluation depends on the older snapshot. Three rules cover the decision:
- Coming from Sonnet 4.6, just move. Same price, higher scores, same 1M context. Pin 4.6 only if you need a frozen model for reproducibility.
- Coming from Opus 4.8, test Sonnet 5 on your hardest real task. If it clears your bar, you drop from $25 to $15 output. If it does not, Opus is still the top tier.
- Run cost-sensitive volume before September. The $2 / $10 intro rate is a real discount with an expiry date. Front-load heavy usage into the window.
Sonnet 5 is a rare kind of release: more capability at the same standard price, with a temporary discount on top. The number to watch is still dollars per finished task on your own workload, not the rate on the card, but for once the rate card moved in the right direction and the sticker did not go up.
Frequently asked questions
- How much does Claude Sonnet 5 cost?
- Standard API pricing is $3 per million input tokens and $15 per million output tokens, the same as Sonnet 4.6. Through August 31, 2026, Anthropic runs introductory pricing of $2 / $10 per million tokens, after which it reverts to the standard rate.
- Is Claude Sonnet 5 more expensive than Sonnet 4.6?
- No. At the standard rate it costs exactly the same, $3 input and $15 output per million tokens. Until August 31, 2026 it is cheaper than Sonnet 4.6 ever was, thanks to the $2 / $10 introductory rate.
- What is the context window and output limit of Claude Sonnet 5?
- Sonnet 5 has a 1M-token context window and a 128k-token maximum output, with adaptive thinking and a January 2026 knowledge cutoff, per Anthropic's models overview. Its API id is claude-sonnet-5.
- How does Claude Sonnet 5 compare to Opus 4.8?
- On SWE-bench Pro, Anthropic's agentic-coding benchmark, Sonnet 5 scores 63.2% against Opus 4.8 at 69.2%, while costing $15 per million output tokens versus Opus at $25. Sonnet 5 buys most of Opus-grade coding for 40% less on output.
- Should I switch from Sonnet 4.6 to Sonnet 5?
- For most workloads, yes. Sonnet 5 is the default, scores higher on every launch benchmark, and costs the same at the standard rate. Keep Sonnet 4.6 pinned only when you need a frozen snapshot for reproducibility.
Sources
- Anthropic. (2026). Introducing Claude Sonnet 5 [vendor launch post; release date, positioning, introductory pricing, launch benchmarks]. Verified 2026-07-01. anthropic.com/news/claude-sonnet-5
- Anthropic. (2026). Models overview [Claude Platform docs; standard pricing $3 / $15, 1M context, 128k output, adaptive thinking, Jan 2026 cutoff,
claude-sonnet-5API id, intro-pricing footnote]. Verified 2026-07-01. platform.claude.com/docs/en/about-claude/models/overview - DataCamp. (2026). Claude Sonnet 5: Features, Benchmarks, and Pricing [secondary; transcription of Anthropic’s launch benchmark table for SWE-bench Pro, Terminal-Bench 2.1, OSWorld-Verified, and Humanity’s Last Exam]. Verified 2026-07-01. datacamp.com/blog/claude-sonnet-5