Claude Sonnet 5 Just Became the Only Frontier Model You Can Actually Buy. Fable Is Dark, GPT-5.6 Sol Is NCD-Gated, Gemini 3.5 Slipped.

On June 30, 2026 Anthropic shipped Claude Sonnet 5 into the public API with a 1M token context window, 85.2 percent on SWE-Bench Verified, 63.2 percent on SWE-Bench Pro, and a $2 in, $10 out introductory rate that holds through August 31 before it steps up to $3 and $15. On paper it reads like an incremental Sonnet tick. It is not. Sonnet 5 shipped into an empty room. Everyone else at the top of the ladder is either dark, gated, or slipped.

Fable 5, Anthropic's own flagship, has been offline since June 12 when Washington issued the export control pull. Mythos 5 went with it. GPT-5.6 Sol is inside a customer-by-customer preview approved by the Office of the National Cyber Director and the Office of Science and Technology Policy, with the broad release still weeks out. Gemini 3.5 Pro missed its June ship date and now targets late July, the second I/O commitment Google has missed on a flagship this year. Meituan open-sourced LongCat-2.0 yesterday, which matters at the price floor, but it is not a US-signed frontier product a Fortune 500 procurement desk will greenlight without a security review.

For the next two to eight weeks, if you want a broadly-available top-tier US model shipping into production traffic today, Sonnet 5 is the list. Not the top of the list. The list.

What Anthropic Actually Shipped

Sonnet 5 is Anthropic's most agentic Sonnet-class model. The company positions it as closing most of the gap to Opus 4.8 on reasoning, tool use, coding, computer use, and long-horizon planning, while landing at Sonnet pricing. The headline specs read like a direct answer to the last six months of enterprise complaints about cost caps.

Context is 1M tokens with adaptive thinking and context compaction, which puts it at parity with GPT-5.5 on window size and closes a gap that had been showing up on tokenized RAG evals. Max output is 128k. Image input is included. Adaptive thinking scales through selectable effort levels up to xhigh, which is Anthropic's answer to OpenAI's max reasoning-effort and ultra modes on Sol.

The pricing lane is the tell. Introductory rate through August 31 is $2 input and $10 output per million tokens, then $3 and $15 at standard rates. There is no long-context premium. Rate limits are higher than 4.6 out of the gate.

One footnote enterprise procurement teams should read twice: the new tokenizer produces roughly 1.0 to 1.35 times more tokens for the same input than the Sonnet 4.6 tokenizer. At standard rates a task that used to cost $X on 4.6 can cost up to 1.35 times as much per unit of text on the new rate card even with a lower headline number. The introductory pricing partially absorbs this. Run a token count on representative traffic before you rewrite your run-rate model.

The Benchmarks

Sonnet 5 leads SWE-Bench Pro, which is the benchmark that matters most for agentic coding right now because it has the least contamination surface. It trails GPT-5.5 on Terminal-Bench 2.1 by three points and would trail GPT-5.6 Sol Ultra by more if Sol were buyable, which it is not. Against the Sonnet 4.6 predecessor the jumps are 12 points on SWE-Bench Pro and 6 points on SWE-Bench Verified, which is a real generational tick and not a fine-tune.

Numbers marked n/d are not directly disclosed by OpenAI in a comparable form. Compare on our benchmarks page where we track third-party reproductions as they come in.

The Empty Room

Sonnet 5 did not launch into a fair fight. It launched into the one broad lane still open. Here is the state of the top of the ladder as of this morning.

Opus 4.8 and GPT-5.5 are still publicly available and both are strong. But the buyable top of the ladder right now is a two-model set: Sonnet 5 at the value tier and Opus 4.8 or GPT-5.5 at the premium tier. Anthropic just owns two of those three positions with the same brand and the same billing surface. Two weeks ago the top of the ladder was Fable 5, Opus 4.8, GPT-5.5, GPT-5.6 Sol (previewing), and Gemini 3.5 Pro (coming). The list shrunk fast.

The Federal Gate Just Handed Anthropic a Distribution Monopoly

The Sonnet 5 release timing is not coincidence. The NCD and OSTP customer-by-customer gate on GPT-5.6 Sol is running two to eight weeks minimum before broad release, based on OpenAI's own posture and the pace of the Fable 5 pull, which is now nineteen days old and still active. That window is exactly the window Sonnet 5 is priced to fill.

There is a real irony here. The same federal release-gating template that pulled Anthropic's own flagship 19 days ago is now the reason the Sonnet 5 launch has the distribution runway it does. The gate that hurts one Anthropic model helps another one. If you were building a procurement strategy at a large enterprise buyer this quarter, the rational move is to standardize agentic workflows on Sonnet 5 now rather than wait on GPT-5.6 Sol's queue position. Every week of production integration on Sonnet 5 during the gate is a week of lock-in that survives whenever the gate lifts.

Anthropic's $47B run-rate context makes this material. The IPO window and confidential S-1 filed June 1 at $965B both need a strong Q3, and Q3 begins today. A two to eight week window where Sonnet 5 is the only broadly-available frontier alternative to Opus 4.8 is exactly the kind of shipping story an S-1 disclosure can lean on. Watch the language in the next amended filing.

What Builders Should Do This Week

If you are running agentic workflows and you have been waiting on GPT-5.6 Sol, stop waiting. Route the workload to Sonnet 5 now. The SWE-Bench Pro delta is real, the Terminal-Bench 2.1 gap versus GPT-5.5 is three points but versus what you can actually buy today it is meaningless. You are picking between Sonnet 5 and Opus 4.8, not Sonnet 5 and Sol.

If you are on Sonnet 4.6, upgrade before August 31. The introductory rate matters, and the SWE-Bench jump is enough to justify the migration even factoring in the tokenizer expansion. Test it against your representative prompt distribution first because the tokenizer difference can eat 20 to 35 percent of the headline savings depending on language, format, and prompt style.

If you are on Opus 4.8, run the cost delta. Sonnet 5 at intro rates is 60 percent cheaper input and output, close on most agentic benchmarks, and behind only on the hardest reasoning tasks. For a big chunk of the traffic currently landing on Opus 4.8, Sonnet 5 is the right default and Opus 4.8 becomes the escalation lane.

If you are on GPT-5.5, the calculus is trickier. GPT-5.5 still wins on Terminal-Bench and on some reasoning benchmarks, and OpenAI's ecosystem and Codex tooling are real. But Sonnet 5 at $2 input beats GPT-5.5 at $5 input by more than 2x on rate card, leads on SWE-Bench Pro, and shares the 1M context window. On pure economics the switch is defensible.

Three Signposts for the Next Ninety Days

One. Does the NCD and OSTP gate on GPT-5.6 Sol lift before August 15? If yes, Sonnet 5's distribution monopoly is a six-week window and Anthropic will have booked one quarter of concentrated demand. If no, Sol's absence from the broad market stretches into Q4 planning and the S-1 amendment language changes.

Two. Does Fable 5 come back? The export-control pull was framed as directive, not permanent. If Fable 5 clears and returns before Q3 ends, Anthropic gets the full three-tier ladder back (Sonnet 5, Opus 4.8, Fable 5) at exactly the moment competitors are still gated. If it does not, Sonnet 5 and Opus 4.8 have to carry the S-1 revenue narrative alone.

Three. Where does Gemini 3.5 Pro actually land in late July? If it ships on time with a competitive SWE-Bench Pro number, the buyable ladder rebalances fast and the empty room closes. If it slips a third time, Google's frontier posture becomes a structural question and the top of the ladder is functionally Anthropic and OpenAI only.

Our Take

Sonnet 5 is a good model. On any other week that would be the whole story. It is not the whole story this week because the release landed inside a distribution vacuum created by a federal gate Anthropic did not build but happens to benefit from asymmetrically. The buyable frontier is a two-model list for the next two to eight weeks, and Anthropic owns both slots. That is not a benchmark win. It is a distribution win. Distribution wins compound faster than benchmarks do.

We are adding Sonnet 5 to our models tracker today and updating the cost calculator with both the introductory rate and the September step-up. We will be watching NCD queue movement on GPT-5.6 Sol closely, because that is the clock on the runway Sonnet 5 got for free.