The Vibe Father
Model Review
Claude Sonnet 5 is the newest member of Anthropic's lineup, released June 30, 2026, and it slots into the family's traditional mid-tier seat: cheaper and faster than Opus, dramatically more capable than Haiku. The pitch writes itself — most of the flagship experience at a fraction of the cost — but pitches are cheap, so we have been running Sonnet 5 through real work and tracking it on our live leaderboard at /benchmarks since day one.
The essentials: $3 per million input tokens, $15 per million output, a 1M-token context window, and 89 tokens per second — the fastest streaming speed of any Claude model above Haiku, and noticeably snappier than Opus 4.8's 60 tok/s. When an agent is iterating in a tight loop, that speed difference is not cosmetic; it compounds across every turn.
The numbers
Here is Sonnet 5 against its family and its closest cross-lab rival. Our Vibe Coding Index weights SWE-bench Verified 40%, Terminal-Bench 30%, LiveCodeBench 30%.
| Model | SWE-bench Verified | Terminal-Bench | LiveCodeBench | Price (in/out per M) |
|---|---|---|---|---|
 Claude Sonnet 5 | 85.2 | not yet published | 82.4 | $3 / $15 |
| Claude Opus 4.8 | 88.6 | 78.9 | 87.8 | $5 / $25 |
 GPT-5.3 Codex | 74.8 | 78.4 | 87.3 | $1.75 / $14 |
One honesty note before anything else: Anthropic has not yet published a Terminal-Bench score for Sonnet 5, so it cannot receive a full Vibe Coding Index score on our board until that lands. We will update the leaderboard the day it does.
On what is published, the headline holds up. Sonnet 5's 85.2 on SWE-bench Verified is 96% of Opus 4.8's 88.6 — on the benchmark that best predicts real repository work, the gap between them is genuinely small. LiveCodeBench is where the mid-tier compromise shows: 82.4 against Opus's 87.8, and notably behind GPT-5.3 Codex's 87.3 despite Codex costing less. If your workload is heavy on fresh algorithmic problem-solving, that row deserves a second look.
In daily use, the speed is what you notice first and the reliability is what you notice over a week. At 89 tok/s, a Sonnet-powered agent turns around a scoped ticket visibly faster than an Opus one, and on routine work — endpoints, components, migrations, test coverage — the output quality is close enough that we frequently cannot tell which model wrote a given diff. Where the gap reappears is ambiguity: hand Sonnet 5 a vaguely specified task and it will pick a reasonable interpretation and sprint, where Opus would have stopped to ask the clarifying question. That is a fine trait in a builder with a good plan in front of it, and a costly one without.
The price math
Our reference heavy month: 50M input tokens, 10M output tokens.
Sonnet 5: 50 × $3 + 10 × $15 = $150 + $150 = $300 per month. The same workload costs $500 on Opus 4.8, $1,000 on Fable 5, and $227.50 on GPT-5.3 Codex.
So the "60% of the price" in our title is literal: $300 against Opus's $500. And the "90% of Opus" is roughly fair on SWE-bench (96%) and slightly generous on LiveCodeBench (94%). For a team burning through agent tokens all day, that $200 monthly gap per heavy seat adds up fast — across five seats it is $12,000 a year. Whether Codex's even-lower $227.50 tempts you instead depends on how much you value Sonnet's 10.4-point SWE-bench lead over Codex; we work through those trades in our cheap coding models guide.
Its best seat on the team
Sonnet 5 is a builder — arguably the builder for teams standardized on Anthropic. In a multi-agent stack it is the model you hand well-scoped implementation tickets: the plan exists, the interfaces are decided, and what you need is fast, correct execution at a price that does not punish volume. Its 85.2 SWE-bench score means it clears the bar for real multi-file repository work, not just snippets, and the 89 tok/s keeps agent loops tight.
It can plan in a pinch, but we would not make that its job — planning quality is where the Opus and Fable 5 premium earns its keep. And with Terminal-Bench unpublished, we cannot yet vouch for it in heavy terminal-driving agentic roles; if that is your workload, GPT-5.5 is the proven pick. The full role-by-role mapping lives in best model for each agent role.
Verdict
Sonnet 5 does what the mid-tier Claude has always done, and does it better than any previous Sonnet: it makes the flagship optional for most work. Paired with an Opus or Fable planner, it is the engine room of a cost-sane multi-agent team — see our best coding model rankings for where it lands overall.
Who should pay for it: high-volume agent operators, Anthropic-standardized teams, and anyone whose token bill on Opus 4.8 makes them wince.
Who should not: if you run a handful of hard tasks a day rather than a firehose of routine ones, the $200 saved over Opus is not worth the reliability gap — buy Opus. If LiveCodeBench-style problem-solving dominates your work, GPT-5.3 Codex scores higher and costs less. And terminal-first agent builders should wait for the missing benchmark or read our cross-lab shootout before committing. For the broad middle of real-world coding, though, Sonnet 5 is exactly the deal it claims to be.