"You now have to justify the flagship. It's no longer the default."

A week after OpenAI claimed its cheaper model matches its flagship, someone checked, and the claim mostly held up.

On Wednesday, The New Stack ran GPT-6.1 Sol against GPT-6 Astra on three real engineering tasks, including CI triage and incident logs. Accuracy tied on every run. Sol was faster on the two longer tests. And it cost about eighteen percent of what Astra did, close to the one-fifth price OpenAI advertises. Artificial Analysis lands in the same place on a broader index. It puts Sol one point below Astra, at less than a quarter of the cost per task.

That's three tasks from one reviewer, plus one benchmark suite. It isn't a verdict on everything. But it's independent, and it agrees with the vendor, which is rarer than it should be.

The same day, the bottom of the market moved too. Anthropic launched Claude Haiku 5.5 at ten cents per million input tokens and fifty cents per million output. That's a ninety percent cut from Haiku 4.5, and it matches OpenAI's GPT-6 Luna on list price. It comes with a one million token context window. Anthropic says real workloads should cost about seventy-five percent less, once token counts change.

Read that last part again, because there are two catches.

First, the cliff. Haiku 5.5's low price only applies to prompts up to one hundred thousand tokens. Above that, it's fifty cents in and two dollars fifty out, five times higher. Anthropic says about ninety percent of Haiku 4.5 requests fell in the short tier. If yours don't, your savings shrink fast.

Second, the tokenizer. The new one counts about thirty percent more tokens for the same text, so the per-token cut overstates the real one. One breakdown puts it at about eighty-seven percent cheaper for short prompts, and around thirty-five percent for long ones.

And OpenAI made its Decisions API public, the typed-answer service from DevDay. Yes, no, or pick one, for a fraction of a full model call.

Put the day together. Mid-tier models now match the flagship on most work. Small models got ten times cheaper. And decision models handle the yes-or-no steps for even less. You now have to justify the flagship. It's no longer the default.

This is the premium tier deflating again, and faster. In September, Ramp's data showed frontier models' share of business token spend falling from a fifty-three percent peak to forty-five. Buyers were already routing down. Now the vendors are pricing for it.

So the work for you is routing, and routing is mostly data work. Log what each request costs and whether it succeeded. Group your traffic by task type. Then run each group on the cheapest tier that passes your own checks. Keep the flagship for the cases the cheaper model measurably fails.

Builder's bench

• Anthropic is giving more security teams access to Claude with fewer cyber restrictions. Meanwhile, Haiku 5.5 ships with stricter cyber safeguards than Haiku 4.5.
• OpenAI published hundreds of AI-generated math results and posted the proofs on GitHub. ChatGPT's new interface swaps walls of text for charts, buttons and mini apps.
• A former OpenAI safety lead says the company ships new capability and risk every Tuesday. Common Sense Media calls ChatGPT for Teens an unacceptable risk. And Simon Willison found OpenAI's rogue agent activity on Wikimedia projects.
• Spanner Omni reached general availability, replacing Google's atomic clocks with software. MLflow 3.17 shipped, InfoQ walked through a session-ordered Kafka pipeline in Go, and OpenSearch veterans launched Infino for agent builders.
• A survey found AI-generated code raises debugging and failure rates and leaves a comprehension gap. COSMIC shut the door on AI code, while GNOME debates AI bug reports. Ben Lorica lists eight security assumptions agents quietly break.
• Microsoft is giving Copilot more control over Windows and your files, and launched new AI PCs. Google made its SynthID detector public and says a hundred eighty billion images and videos now carry its watermark.
• On money, Nous Research hit a one point five billion dollar valuation. Zuckerberg's Biohub is leading a one point eight billion dollar push to model cell behavior, with Google investing. Healthleap raised thirty-eight million dollars.
• Also: Muse is on the iPad, PoeLLM malware is mining crypto on exposed AI servers, and a musician went to prison over ten million dollars in AI-bot streaming fraud.

The through-line: the cheap tier keeps passing tests the expensive tier used to own, so route by your own evidence.

One thing to do today: pull last week's model spend and split it by prompt length. Everything under a hundred thousand tokens is a candidate for a cheaper tier, and that list is your first routing experiment.

Sources covered

1 story · 63 sources (54 news, 2 company, 7 voices)

1. Quoting Ben Affleck
   🎙️Simon Willison↗🎙️Simon Willison↗📰Ars Technica↗📰SiliconANGLE↗📰Ars Technica↗📰The New Stack↗📰The New Stack↗📰TechCrunch↗📰SiliconANGLE↗📰TechCrunch↗📰The Verge↗📰The Decoder↗🎙️Damien Benveniste↗🎙️Simon Willison↗📰The Decoder↗📰The Verge↗📰TechCrunch↗📰The New Stack↗📰TechCrunch↗📰The Verge↗📰TechCrunch↗📰The New Stack↗📰InfoQ↗📰The Decoder↗📰Ars Technica↗📰The Decoder↗📰The Verge↗📰TechCrunch↗📰SiliconANGLE↗🎙️Ben Lorica↗📰The Verge↗📰The Verge↗📰The Decoder↗📰LWN.net↗📰TechCrunch↗📰BleepingComputer↗📰The Decoder↗📰TechCrunch↗📰The Verge↗📰TechCrunch↗📰TechCrunch↗📰The New Stack↗📰The New Stack↗📰Ars Technica↗📰SiliconANGLE↗📰Ars Technica↗📰InfoQ↗📰The Decoder↗🏢MLflow↗📰The Decoder↗🏢Starburst↗📰The Verge↗📰The New Stack↗📰BleepingComputer↗📰The Verge↗📰InfoQ↗📰The Decoder↗🎙️Simon Willison↗📰The Register↗📰InfoQ↗📰SiliconANGLE↗📰InfoQ↗🎙️Simon Willison↗