Claude Sonnet 5 vs Sonnet 4.6: A Skeptic's Migration Guide
Claude Sonnet 5 beats Sonnet 4.6 on SWE-bench Verified (85.2% vs 79.6%) and undercuts it on price until August 31 — a combination rare enough to deserve scrutiny. The catches are real: a new tokenizer that inflates token counts on the same text, breaking API changes around sampling parameters and extended thinking, and a discount that ends September 1. A separate Sonnet 5 rate-limit bucket, distinct from the combined Sonnet 4.x pool, makes running both models during a canary rollout unusually practical. For PMs, the move is an eval-first migration before August 31 — capture the discount window only if your regression suite clears.
Treat this as a dated option, not a mandate. The intro pricing on Claude Sonnet 5 expires August 31, 2026, and a defensible migration — evals, parameter cleanup, tokenizer measurement, canary — takes two to four weeks for most teams. Work backward from that date and the real decision window is now, not mid-August. This week, do one thing: re-run your existing eval suite against claude-sonnet-5 with zero prompt edits, and log three numbers per task type — pass rate, cost per completed task, and token count versus Sonnet 4.6. That last number matters because Anthropic's migration guide documents a tokenizer that maps the same text to more tokens; your effective discount is smaller than the price table implies, and only your traffic can tell you by how much. While the harness is warm, check whether any pipeline sets temperature, top_p, top_k, or manual thinking budgets — those now return 400 errors and represent real engineering work, not config edits. To be fair to Anthropic, the breaking changes are documented, safeguards ship at Opus 4.7/4.8 parity, and the separate Sonnet 5 rate bucket makes canarying unusually cheap. That is a vendor doing migration hygiene right. But documentation isn't validation. If your evals regress, staying on 4.6 costs nothing extra after September 1 — and if they don't, every week you wait is margin left on the table.
Claude Sonnet 5 scores 85.2% on SWE-bench Verified versus 79.6% for Sonnet 4.6, and costs $2/$10 versus $3/$15 until August 31, 2026, per Anthropic's pricing docs.
The API business runs on a dependable rule: the upgrade costs more. Anthropic just broke it. Claude Sonnet 5, released June 30 with the model ID claude-sonnet-5, scores 85.2% on SWE-bench Verified against Sonnet 4.6's 79.6% (llm-stats board, self-reported scores, as of July 2026), and until August 31 it costs $2 per million input tokens and $10 per million output, undercutting 4.6's standing $3/$15.
Better and cheaper, same vendor, same tier. My verdict: migrate, but on your evals' schedule, not Anthropic's promotional one. When a deal looks this clean, the audit is the point.
What's genuinely on the table
Steel-man first, because it's strong.
| Spec | Claude Sonnet 4.6 | Claude Sonnet 5 |
|---|---|---|
| Released | February 17, 2026 | June 30, 2026 |
| Price per 1M tokens (in/out) | $3 / $15 | $2 / $10 through Aug 31; $3 / $15 after |
| SWE-bench Verified | 79.6% | 85.2% |
| Context window | 1M tokens, no long-context premium | 1M tokens |
| Max output | — | 128K (up to 300K via the Batches beta header) |
| Rate-limit bucket | Combined "Sonnet 4.x" (shared with 4.5) | Its own Sonnet 5 bucket |
A 5.6-point jump on SWE-bench Verified is a real gap, not noise, and launch-week reporting put Sonnet 5 at 63.2% on the harder SWE-bench Pro. Availability is boring in the good way: Claude API, AWS Bedrock, Google Vertex, same as 4.6. Safeguards ship at parity with Claude Opus 4.7 and 4.8, per Anthropic, so a compliance review done for the Opus line mostly carries over rather than restarting.
And the discount isn't a stunt with an asterisk. Anthropic's pricing docs state $2/$10 through August 31, standard $3/$15 from September 1. Why a vendor discounts its newest mid-tier model at all is a separate story; our launch coverage argues the IPO calendar explains more than the press release does.
The catches, itemized
Catch one: the discount has a clock, and the clock is doing half the selling. From September 1 the two models cost the same $3/$15, which turns "cheaper and better" into merely "better." The pricing-deadline breakdown covers what that timing means for budgets.
Catch two: the tokenizer. Per Anthropic's migration guide, the same text can map to materially more tokens on Sonnet 5 than on 4.6. That shrinks the 33% sticker cut, and on token-dense workloads it shrinks substantially.
This is a lower per-mile fare on a recalibrated odometer; what you pay depends on both numbers. Only measuring your own traffic settles it.
Catch three: this is not a find-and-replace migration. Per the same migration guide, requests setting temperature, top_p, or top_k to non-default values now return a 400 error, and manual extended-thinking budgets are rejected in favor of a new effort parameter. If your pipeline tunes temperature per task, you have engineering work before you have savings.
Catch four: the headline benchmark is self-reported. That 85.2% is Anthropic's own submission to the llm-stats board. The claim is plausible; independent replication is the missing piece.
Which is exactly why you re-run your own evals. Same-family upgrades have a long history of quiet prompt regressions, and "it's a better Sonnet" is a hypothesis, not a deployment plan.
To be fair to Anthropic, every one of these catches sits in public documentation rather than in someone's production postmortem, and shipping breaking changes loudly beats shipping behavior drift silently.
The rate-limit wrinkle that actually helps
Buried in Anthropic's rate-limit tables is the most useful migration detail nobody markets: Sonnet 5 gets its own bucket — 1,000 requests and 2M input tokens per minute at the Start tier, with 400K output tokens per minute, rising to 5,000 / 5M / 1M at Build — separate from the combined "Sonnet 4.x" bucket that 4.6 shares with 4.5. Two practical consequences follow. Switching doesn't inherit your 4.x traffic pressure, because the new bucket starts empty.
And during migration you can run both models side by side, roughly doubling your effective Sonnet throughput, which makes a proper canary rollout cheaper than it usually is. Add the documented Claude API detail that cached reads don't count toward input-token rate limits, and a well-cached canary barely dents the bucket. The rate-limits reference tracks the current tier tables as they change.
Migrate like you don't trust it
The checklist, in order:
- Re-run your eval suite against
claude-sonnet-5before production sees a single request. Same prompts, same graders, no edits. - Strip sampling parameters and port any manual thinking budgets to the effort parameter; confirm nothing returns a 400.
- Measure real token counts on a sample of production traffic to price the tokenizer change against the discount.
- Canary 5–10% of traffic in the separate bucket, and compare cost per completed task, not cost per token.
- Benchmark a rival in the same pass. The harness is already warm, and our GPT-5.6 Terra comparison lays out the closest cross-vendor alternative at this price point.
Staying on 4.6 is defensible in exactly three cases: your evals regress, your workload depends on the sampling controls Sonnet 5 removed, or your team has no eval capacity this quarter. After September 1, price parity means staying is a performance decision, not a financial one. The live model tracker has current numbers, because this paragraph will age.
The harder question isn't whether Sonnet 5 beats 4.6. On the available evidence, it does. It's whether a two-month discount should compress your migration timeline at all.
Anthropic has now demonstrated it can move thousands of engineering roadmaps with a pricing footnote and a date. Before August 31, decide which one sets your cadence: your eval results, or a vendor's promotional calendar on the way to an IPO.
Get this in your inbox. AI Rundown Daily delivers original briefings every morning — free. Subscribe →
Frequently Asked Questions
The headline numbers are real: 85.2% versus 79.6% on SWE-bench Verified per the llm-stats board (self-reported, as of July 2026), and $2/$10 versus $3/$15 per Anthropic's pricing docs. The catches are the fine print. The discount ends August 31, 2026, after which both models cost $3/$15, and Anthropic's migration guide documents a new tokenizer that produces more tokens for the same text, which erodes part of the savings. The benchmark gain is also Anthropic's own submission, so validate it against your workload before trusting it.
Two documented hard failures, per Anthropic's migration guide: requests that set temperature, top_p, or top_k to non-default values return a 400 error, and manual extended-thinking budgets are rejected in favor of a new effort parameter. Beyond the hard failures, the tokenizer change alters token counts and therefore cost per request, and prompt behavior can shift even within the same model family. That is why the sane sequence is evals first, parameter cleanup second, a 5–10% canary third — using Sonnet 5's separate rate-limit bucket so the rollout doesn't compete with your existing Sonnet 4.x traffic.
Three defensible cases: your eval suite regresses on Sonnet 5, your workload depends on the sampling controls (temperature, top_p, top_k) that Sonnet 5 removed, or your team lacks eval capacity this quarter and can't migrate responsibly. From September 1, 2026 the two models cost the same $3/$15, so staying carries no financial penalty — only a 5.6-point benchmark gap. The longer-term risk is roadmap drift: vendor attention, tooling, and future improvements will concentrate on the Sonnet 5 line, so treat staying as a deferral with a revisit date, not a permanent decision.
