Good morning. Today’s briefing is dominated by model drops with real substance behind them: Anthropic finally fixed Haiku’s pricing problem, OpenAI’s math results are getting serious endorsement from actual mathematicians (including a measured response from Terence Tao), and GPT-6 arrived with a UI makeover that not everyone wants. On the business side, Meta and Microsoft are quietly weaning employees off Claude, and Zuckerberg is pulling Google and Meta into a $1.8B bet on a “virtual cell.”
Claude Haiku 5.5 lands at Luna pricing. Anthropic’s new small model is roughly 90% cheaper than Haiku 4.5 for prompts under 100K tokens, matching GPT-6 Luna’s price exactly at that tier, and it introduces adjustable effort levels (Low through Max) to the Haiku line for the first time. The announcement also quietly includes a 50% cut to Sonnet 5.5 cache reads and new monthly API credits for Max and Team subscribers — $100 to $500 depending on plan. The HN crowd mostly cheered, though several flagged the 100K-token pricing cliff (5x markup above that threshold) as awkward for cost planning, and Plotly’s benchmark team reported their data analytics score jumped two letter grades at 9x lower cost.
GPT-6 arrives with “Intelligent UI” — and ads concerns. OpenAI announced GPT-6 for everyone alongside a feature that renders responses as checklists, cards, and interactive layouts rather than plain text. The system card acknowledges statistically significant regressions on self-harm and extremism evaluations versus GPT-5.6. HN reaction was prickly: one commenter said the heavily formatted output feels “condescending” and “treats you like a child,” while several pointed to OpenAI’s new in-chat advertising format announced days ago as the likely motivation for all that visual real estate.
The math drop keeps getting bigger. Following yesterday’s dump of 722 manuscripts, specialists are confirming the results are the real thing. OpenAI’s post highlights a proof of the Unique Games Conjecture — a load-bearing assumption in complexity theory — alongside Barnette’s Conjecture and progress on four of the seven Millennium Prize problems. One HN commenter who spent 24 years thinking about Barnette’s shared a bittersweet note about seeing it solved. Terence Tao’s measured response acknowledges AI will transform mathematics but pushes back on the practice of dumping proofs without exposition, verification, or corollaries — leaving human mathematicians to do the janitorial work. He suggests the field may need to reorient toward simpler proofs and better intuitions rather than solution-hunting.
Mistral Large 4 lands with EU sovereignty subtext. We covered the headline numbers yesterday, but the official post is now up with full benchmarks. The vision results (42% on Dense 200, edging GPT-6 Astra) and cybersecurity scores (82% on CyberGym-E2E) are the real highlights; broader reasoning still trails the frontier by a few points on Vals Index. One HN commenter captured the ambivalence well: good to see competition and a non-US/China lab in the race, but “cheering for the last kid crossing the finish line” isn’t quite the same as cheering for a winner.
Meta and Microsoft cut Claude usage. Both companies are pushing employees onto in-house tools — Microsoft toward Copilot and OpenAI frameworks, Meta toward MetaCode and Muse Code — per RS Web Solutions. Microsoft slashed per-employee monthly AI budgets from $100,000 to around $10,000, and Meta’s Claude Code user count halved from 60K to 30K. Meta still spent over $105M on Claude Code in a 28-day window, so the drawdown is relative. HN consensus: this is dogfooding and data governance more than a quality signal, but one commenter noted Anthropic reportedly gets a quarter of its revenue from two clients, and Meta is probably one of them.
Zuckerberg’s “virtual cell” draws Google and Meta money. Google DeepMind, Meta, and Isomorphic Labs are putting $300M into Biohub as part of a $1.8B “Virtual Biology” initiative aimed at building an AI model of a working cell. The US DOE is adding $500M and NIH is contributing $500M in datasets. The pitch is running biological experiments digitally; the subtext is a rare moment of DeepMind and Meta collaborating rather than competing.
Quick hits. Nous Research confirmed a $90M Series B at a $1.5B valuation and launched “Hermes for Businesses”; the open-source Hermes model reportedly accounts for ~2.5% of global AI token usage. Docker released Docker Agent, a YAML-based agent orchestration CLI that ships with Docker Desktop 4.63+ — HN reception was lukewarm, with one commenter noting agent harnesses are “turning into JS frameworks from yesteryear.” And Google launched SynthID Detector for identifying watermarked AI content; one user’s informal test caught 6 of 8 AI images, though the auth requirement and lack of bulk classification drew complaints.
That’s the morning. Watch Haiku 5.5 adoption numbers — if the pricing really does match Luna’s at the volume tier most people care about, it’s the first time in a while Anthropic has had a genuine cost argument to make.