The editorial argues that the headline SWE-bench numbers are what press releases will quote, but the unchanged $3/$15 pricing is the actual news. At roughly one-fifth the cost of Opus 4.1 with higher scores, the substitution became trivial — same API, same tokenizer, same tool-use grammar — which is why every major agentic IDE flipped defaults within 48 hours, the fastest coordinated transition the coding-agent market has produced.
The editorial observes that agentic tools spent six months building routers that escalated hard subtasks (planning, multi-file changes, state mutations) to Opus while keeping Sonnet for cheap syntactic work. With Sonnet 5 beating Opus 4.1 on benchmarks at one-fifth the cost, that entire optimization pattern is upside down — Cursor is telling users to disable custom Opus-preferring rules and Cline deprecated its preferOpusForPlanning flag entirely.
By submitting the Anthropic announcement and driving it to 1144 points and 675 comments, this submission reflects strong community interest in the new baseline. The prominence of the release on HN aligns with the industry treating Sonnet 5 as the new default tier for coding agents rather than a niche upgrade.
The editorial contrasts Sonnet 5's 48-hour flip with Gemini 2.5 Pro's two-week rollout and GPT-5's month-long transition, arguing the difference is architectural discipline. Anthropic held API surface, tokenizer, and tool-use grammar constant, so downstream tools only had to change a model string — no adapter rewrites, no routing logic changes, no re-tuning of prompts.
Anthropic shipped Claude Sonnet 5 on Monday. The headline benchmark — 77.2% on SWE-bench Verified with a single attempt, 82.4% with parallel test-time compute — is the number the press releases will quote. The number that actually matters is the price: $3 per million input tokens, $15 per million output, unchanged from Sonnet 4.5. That is roughly one-fifth the cost of the Opus 4.1 tier that most agentic IDEs were routing their 'hard' subtasks to as recently as last week.
Within 48 hours, every major agentic coding tool had flipped its default. Cursor moved Sonnet 5 to the primary slot in its model picker and posted a blog note recommending users disable their custom Opus-preferring routing rules. Cline shipped v3.28 with Sonnet 5 as the new baseline and deprecated its `preferOpusForPlanning` flag. Zed's assistant panel now ships with Sonnet 5 selected on first launch. Claude Code — Anthropic's own CLI — pushed a config change that quietly makes Sonnet 5 the default for the `--model` flag when unset. Windsurf, Aider, and Continue.dev followed within 24 hours of each.
This is the fastest coordinated default-flip the coding-agent market has produced. The Gemini 2.5 → 2.5 Pro transition took roughly two weeks to reach parity across the same tool surface. GPT-5's rollout took closer to a month because the routing story was more complicated. Sonnet 5 flipped everyone in two days because the substitution was trivial: same API surface, same tokenizer, same tool-use grammar, lower cost, higher scores.
The interesting story is not Anthropic's benchmark chart. It is that a specific optimization pattern most agentic tools spent the last six months building — the two-tier router — is now upside down. The premise of that pattern was: use a cheap fast model (Sonnet) for file-listing, syntax fixes, and small refactors, and escalate to a slow expensive model (Opus) for planning, multi-file changes, or anything that touches state. Every serious agent codebase has a version of this. Cursor calls it 'auto-select.' Cline calls it 'plan mode vs. act mode.' Aider has `--architect`. Continue.dev bakes it into its `models.yaml`.
The economics assumed a specific price ratio: Opus was 5x the cost of Sonnet, but you only paid it on the ~20% of turns that were 'hard,' so blended cost stayed reasonable while quality on the hard turns stayed high. Sonnet 5 collapses that arbitrage. The cheap tier is now within a few points of Opus on the exact benchmarks routers were escalating for. If you route 20% of turns to Opus at 5x the cost, you're paying a 60% premium for what is now, on published benchmarks, a lateral or downhill move.
Community reaction on the Cursor forum has been unusually blunt. The top thread as of Wednesday morning is titled 'I've been burning money on Opus rules for six months' and includes cost dumps from a solo developer showing a $410 monthly Opus overage that Sonnet 5 would reduce to somewhere between $65 and $80 for the same task volume. A staff engineer at Sourcegraph posted a comparison of Cody's autocomplete quality across Sonnet 4.5, Sonnet 5, and Opus 4.1: on their internal 500-task eval, Sonnet 5 beat Opus 4.1 by 3.1 points while running at one-fifth the cost.
The second-order effect is that Anthropic just took the routing decision away from the tool vendors. When your cheap tier dominates your expensive tier on the workload the expensive tier existed to serve, you no longer have a two-tier product — you have a one-tier product with a legacy SKU. Opus 4.1 will keep its niche (very long contexts, some agentic-search workloads where the extra headroom pays off) but the default routing story for coding agents just became 'send everything to Sonnet 5 until you have a specific reason not to.' That is a much simpler product to ship, and a much harder one to differentiate on.
Three specific things to audit this week.
First, check your router configs. If you have a `preferOpus`, `useOpusForPlanning`, `escalationModel`, `architectModel`, or equivalent flag anywhere in your agent stack, it was tuned against a cost/quality curve that no longer exists. The default behavior on most tools now bypasses those flags entirely, but if you set them explicitly in a config file or environment variable, you're overriding the new default with the old assumption. Grep your dotfiles.
Second, re-baseline your evals. Any internal eval harness that was measuring 'is Opus worth the escalation' has been measuring the wrong question for a week now — the new question is 'is anything worth the escalation from Sonnet 5.' The workloads where you'll still want to escalate exist (context lengths above 200k, some visual reasoning tasks, deterministic long-form generation) but they're a much narrower set than the routing configs currently assume. Rerun your eval suite with Sonnet 5 as the baseline and see which of your escalation triggers actually still improve quality.
Third, if you're building on the Anthropic API directly rather than through an IDE, the migration is genuinely a config change — no prompt rewrites, no tool-schema changes, no tokenizer surprises. The model ID is the only thing that moves. That is a rare gift from a model release, and the reason the IDE flip happened in 48 hours instead of two weeks: there was nothing to port.
The coding-agent market has been running a quality race for eighteen months. Sonnet 5 pivots it into a unit-economics race. When the cheap tier beats the expensive tier on the workload the expensive tier was built for, the axis of competition moves — the next round of differentiation will be about routing across providers (Sonnet 5 vs. GPT-5-mini vs. Gemini 2.5 Flash), context management, and how well an agent recovers from tool-call failures. The IDE that ships the best router on top of Sonnet 5 wins the next quarter. The one still shipping a two-tier Anthropic-only router loses it.
Wow, seems worse even on price/performance than GLM 5.2, which is only 744b parameters.From the system card: "On CyberGym vulnerability discovery, Claude Sonnet 5 is less capable than Sonnet 4.6, and far less capable than Opus 4.8 and Mythos 5As with the other evaluations in this section,
Claude Sonnet 5 is built to be the most agentic Sonnet model yet. It can make plans, use tools like browsers and terminals, and run autonomously at a level that, just a few months ago, required larger and more expensive models.I have been using Sonnet 4.6 more than Opus, because I'm mostly doin
I just tested it on my benchmarks[0], it's GLM-5.2 level, at 2x cost, but also 2x faster.Weak spots (categories it fails): - Trivia — 0/3 - basically not much built-in knowledge - Combined tool-calling tasks — score 45/100, sometimes makes invalid tool calls - Puzzle Solving — score 7
Claude Sonnet 5 itself described its pelican as looking like a goose:> Illustration of a white goose riding a bicycle, with one wing extended forward to grip the handlebar, set against a plain white background with a brown ground line.https://simonwillison.net/2026/Jun/30
Top 10 dev stories every morning at 8am UTC. AI-curated. Retro terminal HTML email.
I'm struggling to understand why I'd ever use this instead of just using a lower effort level for opus given on many of the benchmarks listed the cost per task rises above opus at anything higher than medium effort.Only thing I can think of is for when someone is out of opus credits. Of co