What Happened
On June 30, 2026, Anthropic released Claude Sonnet 5, calling it the most agentic Sonnet-class model it has ever shipped. It became the default model for every Free and Pro user on Claude.ai starting July 1, and it's also live in Claude Code, the Claude Platform, and third-party tools like Cursor, GitHub Copilot, and VS Code. The launch lands in the middle of a genuinely wild few weeks for Anthropic: Claude Fable 5 had just been pulled offline on June 12 under U.S. export controls and only came back online on July 1, the same week Sonnet 5 shipped.
Sonnet 5 isn't just an incremental update. Anthropic's own numbers show it closing most of the gap with Opus 4.8 on agentic tasks — planning, tool use, coding, browsing — while costing a fraction as much. That puts real pressure on both Opus 4.8 (Anthropic's own flagship) and GPT-5.5 (OpenAI's flagship, released back in April), because for the first time, a mid-tier model is genuinely competitive with the top tier.
Background: Where Each Model Sits
Claude Sonnet 5 is the sixth generation of Anthropic's mid-tier model, sitting between Haiku 4.5 and the Opus line. Historically, Sonnet models were the ones that first made "agentic" coding assistants feel real — Sonnet 3.5, 3.6, and 3.7 all built that reputation. Sonnet 5 is Anthropic's attempt to bring Opus-level agentic performance down to Sonnet pricing.
Claude Opus 4.8 remains Anthropic's most capable generally available model (Fable 5 and Mythos 5 sit above it but are far more restricted). Opus is still the model Anthropic recommends for the hardest reasoning tasks and any work that touches cybersecurity, where Sonnet 5 is deliberately weaker by design.
GPT-5.5 is OpenAI's flagship, released in April 2026 as a response to Anthropic's Opus 4.7. It was pitched around agentic coding, computer use, and long-horizon knowledge work, with a notably faster release cadence than OpenAI's older roughly-annual model cycle.
Why It Matters
For most of 2026, if you wanted the best agentic performance from any lab, you paid flagship prices. Sonnet 5 changes that math for Anthropic's lineup specifically. On several benchmarks it's within a few points of Opus 4.8, and on one — Terminal-Bench 2.1, which tests command-line and multi-step tool orchestration — it actually beats Opus 4.8 outright. That's the headline: the "good enough" tier just got a lot more capable, which is exactly the kind of shift that forces competitors to respond on price, not just capability.
Industry Impact
The pattern across all three labs right now is the same: ship "agentic" as the headline feature, then compete on how affordably that agentic work can run. OpenAI's GPT-5.6 Sol, previewed just before Sonnet 5's launch, leans on the same pitch — splitting work across subagents for longer autonomous runs. Google's Gemini 3.5 Flash made an identical argument back in May. Agentic capability is no longer a differentiator on its own; it's table stakes. The real competition now is over cost-per-completed-task, not cost-per-token, which is a much harder number for any company to advertise cleanly.
That's also why Sonnet 5 ships with a pricing footnote worth taking seriously. Anthropic changed Sonnet 5's tokenizer, and the same text can now map to up to 1.35x more tokens than it did under Sonnet 4.6. Anthropic set the introductory price low enough to make the switch "roughly cost-neutral" through August 31 — but standard pricing after that date ($3/$15 per million tokens) is identical to what Sonnet 4.6 charged, while the token count for the same workload is higher. Independent analysis from Artificial Analysis found an average task on Sonnet 5 at standard pricing costs about $2.29 to complete, versus roughly $1.20 for Sonnet 4.6 — nearly double, despite the "flat" rate card.
Consumer Impact
If you're a Claude.ai Free or Pro subscriber, none of this pricing complexity touches you directly — Sonnet 5 is simply your new default model as of July 1, at no extra cost, and it's a meaningfully better assistant for anything involving multi-step tasks, tool use, or coding help. If you're a developer paying per token through the API, the calculus is more interesting: Sonnet 5 at introductory pricing is a clear upgrade over sticking with Sonnet 4.6 or paying Opus 4.8 rates, but it's worth testing your actual workloads before September 1, when both the price step-up and the tokenizer change land at the same time.
Expert Analysis
Early tester feedback across all three models converges on the same theme: these updates are less about raw intelligence gains and more about "finishing the job." Testers of Sonnet 5 described it completing multi-step work — updating records, then sending a follow-up announcement — that older Sonnet versions would abandon partway through. OpenAI made nearly the identical claim about GPT-5.5, saying it "keeps going until it is done" rather than stopping after a plausible first answer.
On raw benchmark numbers, no single model wins across the board, and the benchmarks themselves aren't always directly comparable across labs since methodologies differ. Sonnet 5 scores 63.2% on agentic coding (SWE-bench Pro) against Opus 4.8's 69.2%, but flips the result on Terminal-Bench 2.1, where it hits 80.4% versus Opus 4.8's 74.6%. GPT-5.5 posts 82.7% on the earlier Terminal-Bench 2.0 version and leads on GDPval knowledge-work testing at 84.9%. None of these numbers alone tells you which model to use — they tell you which model was tuned for which kind of task.
Pricing and Benchmark Comparison
| Model | Input / Output (per 1M tokens) | Context Window | Agentic Coding | Best For |
|---|---|---|---|---|
| Claude Sonnet 5 (intro, thru Aug 31) | $2 / $10 | 1M tokens | 63.2% (SWE-bench Pro) | Everyday coding, agents on a budget |
| Claude Sonnet 5 (standard, from Sep 1) | $3 / $15 | 1M tokens | 63.2% (SWE-bench Pro) | Same, at post-promo pricing |
| Claude Opus 4.8 | $5 / $25 | — | 69.2% (SWE-bench Pro) | Hardest reasoning, cybersecurity, high-stakes work |
| GPT-5.5 | $5 / $30 | 1.05M tokens | 58.6% (SWE-bench Pro) | Knowledge work, long-context research |
| GPT-5.5 Pro | $30 / $180 | 1.05M tokens | Highest accuracy tier | Research-grade, long-horizon, high-stakes tasks only |
Note: benchmark methodologies differ slightly between labs, so treat these as directional rather than perfectly apples-to-apples.
Timeline
| Date | Event |
|---|---|
| April 23, 2026 | OpenAI releases GPT-5.5 and GPT-5.5 Pro |
| June 12, 2026 | U.S. export controls suspend Claude Fable 5 and Mythos 5 |
| June 30, 2026 | Anthropic releases Claude Sonnet 5; export controls on Fable 5 lifted the same day |
| July 1, 2026 | Sonnet 5 becomes default for Free/Pro users; Fable 5 access restored |
| August 31, 2026 | Sonnet 5 introductory pricing ends |
| September 1, 2026 | Sonnet 5 standard pricing ($3/$15) and new tokenizer both take full effect |
FAQ
Is Claude Sonnet 5 actually cheaper than Opus 4.8?
Yes, at both introductory and standard pricing. Standard Sonnet 5 ($3/$15) is roughly 40% cheaper than Opus 4.8 ($5/$25) per million tokens, and during the introductory window through August 31 the gap is even wider.
Does Sonnet 5 replace Opus 4.8?
Not entirely. Opus 4.8 still leads on the hardest reasoning benchmarks and is the model Anthropic recommends for cybersecurity-adjacent work, where Sonnet 5 was deliberately not trained to perform well.
Why did Sonnet 5's price go up on September 1 if the rate card looks flat?
The sticker price after August 31 ($3/$15) matches what Sonnet 4.6 charged, but Sonnet 5's new tokenizer can turn the same text into up to 35% more tokens. The introductory pricing was set specifically to offset that increase temporarily.
How does GPT-5.5 compare on price to Sonnet 5 and Opus 4.8?
GPT-5.5 ($5/$30) costs the same on input as Opus 4.8 but more on output, and is meaningfully more expensive than Sonnet 5 at either pricing tier.
Which model should I actually use?
For budget-conscious coding agents and everyday assistant work, Sonnet 5 is the strongest value right now. For the hardest reasoning or any cybersecurity-adjacent task, Opus 4.8. For long-context knowledge work and research-heavy tasks, GPT-5.5 is a strong option, with GPT-5.5 Pro reserved for the highest-stakes work only.
My Take
[Personalize this section with your own take before publishing — a few angles to consider: the fact that Sonnet 5's "cost-neutral" framing has a September expiration date feels like it deserves more scrutiny than it's getting; or, as someone who builds front-end tools, note which of these you'd actually reach for in a Cursor/Copilot workflow and why.]
Key Takeaways
- Claude Sonnet 5 closes most of the agentic-performance gap with Opus 4.8 at roughly 40-60% of the cost.
- Sonnet 5's pricing is only "cost-neutral" through August 31 — after that, a tokenizer change and a price increase land at the same time.
- Opus 4.8 still leads on the hardest reasoning tasks and remains the recommended model for cybersecurity-adjacent work.
- GPT-5.5 leads on knowledge-work benchmarks like GDPval and offers the largest context window of the three.
- None of the three models wins every benchmark — pick based on your actual workload, not a single leaderboard number.
