Claude 4.6 Sonnet shipped on March 10, 2026, at $3 input and $15 output per million tokens, with a 200,000-token context window and 77.5% on SWE-bench Verified. Thirty-five days later, Anthropic released Opus 4.7 at 82.4% on the same benchmark. Four days after that, Google put Gemini 3.1 Pro on the board with a 2M-token window and a 94.3% GPQA Diamond score. A week later, GPT-5.5 landed at 80.8% SWE-bench with native audio. The workhorse tier Sonnet defined in March is not the same product it was when the pricing was set.
Where Sonnet 4.6 Landed on March 10, and What Shipped After It
Let me walk you through the calendar, because the calendar is the story here. On March 10, Sonnet 4.6 was the second-best coding model on the SWE-bench Verified board — behind only its own predecessor tier, GPT-5.4, which had landed five days earlier at the same $3/$15 price point. Anthropic priced Sonnet to match the incumbent flagship's number-for-number curve, then delivered a 4.6-point SWE-bench lead. For thirty-five days, that positioning held.
Then April 15 arrived. Opus 4.7 posted 82.4% on SWE-bench Verified, 91.2% on GPQA Diamond, 94.0% on HumanEval, and a 1M context window — for $15 input and $75 output. Three days later, Google's Gemini 3.1 Pro shipped at $2.50/$10 with a 2M context window and the highest GPQA Diamond score anyone has posted publicly. Then GPT-5.5 on April 22 at $5/$25, 80.8% SWE-bench, and full audio modality on top of vision.
Here is what nobody in the launch-week posts is willing to say out loud: Sonnet 4.6 did not get worse in April. It got surrounded. That is a different problem, and the fix is a different decision.
The 77.5 SWE-bench Number Is the Only One That Matters Here
We keep coming back to SWE-bench Verified because it is the only benchmark on the March-to-April release cycle where the delta actually maps to a workload people are paying for at scale. GPQA Diamond tells you which model can pass a PhD-level physics quiz. HumanEval measures function-completion on toy problems. MMLU is saturated. SWE-bench Verified — the human-audited subset of real GitHub issues resolved by real patches — is what maps to "can this model close a Jira ticket without a human rewriting half the diff."
Here is the April 22 board:
Claude 4.7 Opus: 82.4%. GPT-5.5: 80.8%. Claude 4.6 Sonnet: 77.5%. Gemini 3.1 Pro: 75.3%. Grok 4: 74.0%.
Sonnet 4.6 is now the third-best coding model in the world, one tier below its own family flagship, two behind the OpenAI April release. Methodology note: SWE-bench Verified is agentic — the score reflects the model plus its harness plus its tool-use scaffolding. Anthropic's reported number uses their internal harness. OpenAI reports on a different one. The 1.6-point delta between Opus 4.7 and GPT-5.5 is inside harness-noise. The 4.9-point delta between Sonnet 4.6 and Opus 4.7 is not.
If your work is dominated by SWE-bench-shaped tasks — repo-scoped bug fixes, multi-file refactors, agentic edits — the 77.5 is a real, structural cost you eat every time you route to Sonnet instead of the tier above it.
Pricing: $3/$15 Held the Line, But the Line Moved
Here is where the receipts get interesting.
Claude 4.6 Sonnet: $3.00 input, $15.00 output per million tokens. GPT-5.4 (March 5 flagship): $3.00 / $15.00 — matched. GPT-5.5 (April 22 flagship): $5.00 / $25.00 — a 67% jump for the successor tier. Gemini 3.1 Pro (April 18): $2.50 / $10.00 — undercut Sonnet by 33% on output. Opus 4.7 (April 15): $15.00 / $75.00 — held flat from Opus 4.
The $3/$15 slot Anthropic staked out in March is now a crowded corner. Gemini 3.1 Pro is priced 33% lower on output for a model that beats Sonnet on GPQA (94.3 vs no comparable public Sonnet number in this grounding) and matches it within a few points on SWE-bench. Grok 4 sits at $3/$15 flat with a 74.0% SWE-bench — the same slot, worse on the code number. And the tier above — GPT-5.5 at $5/$25 — is not a 5x jump in cost like the Opus tier, it is a 67% jump for a 3.3-point SWE-bench lift.
The primary-document cross-reference worth unwinding: Anthropic's March 10 announcement described Sonnet 4.6 as "our workhorse tier with prompt caching by default." Anthropic's April 15 announcement described Opus 4.7 as "frontier reasoning and coding at the 1M-context tier." Both are technically operative. The first was written when Sonnet was the highest-performing model in its price band. The second was written when it stopped being that. Prompt caching by default is still real. It is what keeps Sonnet on the table at all.
The 200K Context Window Is Now the Ceiling, Not the Floor
Here is a number that stopped being competitive faster than the price did: 200,000 tokens.
In March, 200K was the standard workhorse ceiling. Every model in Sonnet's price band shipped with 128K–200K. In April: Gemini 3.1 Pro at 2,000,000 tokens. Opus 4.7 at 1,000,000. GPT-5.5 at 400,000. Gemini 3 Flash (the February release, priced at $0.30/$1.20) at 1,000,000. The $0.30-per-million tier now has 5x the context window of the $3-per-million tier.
Think about what that actually means for a workload. If you are chunking a monorepo to fit into 200K, you are engineering around a limit that no other frontier model in April requires you to engineer around. Retrieval scaffolding, chunk-and-map orchestration, sliding-window summarization — this is the infrastructure tax Sonnet 4.6 users are now paying that Gemini 3.1 Pro users are not.
The workaround is not free. Multi-turn context reassembly on 200K adds latency and adds tokens (because you are re-uploading chunks), and neither cost was priced into the $3/$15 assumption when Sonnet launched. There is a case where 200K is enough — most agentic subtask contexts fit — but the ceiling used to be aspirational. Now it is a constraint you can feel.
Who Should Actually Still Be Reaching for Sonnet 4.6 in Late April
Alright. Here is the honest read.
Sonnet 4.6 is still the right call for three specific workloads. First: agentic scaffolds where prompt caching is doing structural work — long system prompts, repeated tool definitions, sub-agent orchestration where the same context is hit hundreds of times per session. Anthropic's caching pricing is the industry's most aggressive on cache-hit, and if your workload has hit rates above 40%, the effective per-token cost drops below what Gemini 3.1 Pro can match at list price. This is the case Claude Code is built on and it is the case Cursor's default routing to Sonnet reflects.
Second: shops that have already invested in the Anthropic prompt idiom — constitutional-style system prompts, XML tag conventions, the tool-use format — and where retooling for a Google or OpenAI SDK is measured in engineering weeks. Migration cost is real. Sonnet 4.6 is the price of not migrating this quarter.
Third: workloads where 77.5% SWE-bench is enough. Not every ticket needs the top of the leaderboard. If your workflow has a human reviewer in the loop and the model is drafting, the 4.9-point delta to Opus 4.7 is a cost-benefit calculation, not a correctness argument. At 5x the output pricing for Opus, you eat a lot of Sonnet retries before Opus pays back.
The reader this piece is for should be running the numbers on their own request distribution before the next billing cycle closes. The models above Sonnet on the leaderboard are not experiments anymore. They are shipped, priced, and taking share.
None of this tells you what happens on the next Anthropic release — whether Sonnet 4.7 lands at the same $3/$15 slot with a 400K context and closes the gap, or whether Anthropic moves the workhorse tier up-market to $5/$20 and cedes the low end to Gemini. That decision is where the next quarter's routing math actually gets decided, and it is not one this piece can answer.
FAQ
Is Claude 4.6 Sonnet still worth using after the April 2026 frontier releases?
For workloads that lean on prompt caching, existing Anthropic tooling, or 200K-token contexts where the ceiling is comfortable, yes. Sonnet 4.6 remains the cheapest Anthropic model with the current-generation reasoning stack, and its 77.5% SWE-bench Verified score is competent for most agentic subtasks. For frontier coding work — repo-wide agentic edits with tight correctness budgets — Opus 4.7 or GPT-5.5 now sit clearly ahead, and the routing decision should reflect that.
How does Claude 4.6 Sonnet's pricing compare to Gemini 3.1 Pro?
Sonnet 4.6 is priced at $3 input and $15 output per million tokens. Gemini 3.1 Pro, released April 18, 2026, is priced at $2.50 input and $10 output — a 33% cheaper output tier for a model with a 2M context window versus Sonnet's 200K. On SWE-bench Verified, Sonnet 4.6 still leads Gemini 3.1 Pro (77.5 vs 75.3), but on GPQA Diamond, Gemini posts 94.3%, the highest publicly reported. The two models are targeting different failure modes.
What is the SWE-bench Verified difference between Sonnet 4.6 and Opus 4.7?
Sonnet 4.6 scored 77.5% on SWE-bench Verified. Opus 4.7 scored 82.4%. The 4.9-point delta is above harness-noise and represents a real capability gap on repo-scoped software engineering tasks. The pricing gap between the two tiers is 5x on both input ($3 vs $15) and output ($15 vs $75), so the correct decision depends on how often your workload benefits from the lift. High-stakes, low-review-budget tasks lean Opus. Draft-and-review workflows can still route to Sonnet.
Does Claude 4.6 Sonnet support prompt caching by default?
Yes. Prompt caching is enabled by default on Sonnet 4.6, per the March 10, 2026 release notes. This is the single most important economic feature the model retains against the April challengers. Workloads with high cache-hit rates — long system prompts, repeated tool definitions, agentic scaffolds — see effective per-token costs drop below the list price, which is what makes Sonnet 4.6 still competitive against Gemini 3.1 Pro's lower list pricing for many production deployments.
Which coding tools currently default to Claude 4.6 Sonnet?
Windsurf, from Codeium, currently lists Claude 4.6 Sonnet as its primary model at $15/month. Cursor's primary listed model is Claude 4.7 Opus at $20/month, though its routing exposes Sonnet as a lower-cost option. Claude Code — Anthropic's own CLI agent — is bundled into Claude Pro at $20/month and defaults to the Opus tier. If you are on Windsurf, Sonnet 4.6 is what you are using unless you explicitly route elsewhere.
Why did Anthropic hold Opus 4.7 pricing flat at $15/$75?
The Opus 4.7 pricing held at $15 input and $75 output per million tokens — identical to the previous Opus tier. This is unusual for a frontier release that posts a capability jump on SWE-bench Verified, HumanEval, and GPQA Diamond. The category context: GPT-5.5 raised the flagship OpenAI tier from $3/$15 to $5/$25, a 67% jump, while Gemini 3.1 Pro undercut on output. Anthropic's flat-pricing posture on Opus is a bet that customers value price stability over the capability-priced ladder OpenAI is climbing.