Thomas Sottiaux, OpenAI's Codex lead, publicly told developers this week to route Claude Code at gpt-5.6-sol via CLIProxyAPI. The same day, mweinbach confirmed he'd been running GPT models inside Claude Code for a while. Meanwhile Anthropic extended Fable 5 access and kept Claude Code rate limits 50% higher — not to defend the model, but to defend the harness. The gap between those two moves is the whole story.
What's actually happening is a split in where value is being staked. OpenAI is betting the model is the callable — harness is just plumbing, so endorsing a competitor's runtime costs nothing. Anthropic is betting the harness is the runtime — model is the swappable part, so the subscription that keeps developers inside Claude Code is worth subsidizing. Pietro Schirano's Coding Agent Index already shows Fable's price-per-task premium evaporating at the API layer. If the harness wins and the model is a commodity, Anthropic's bet compounds. If the model wins and harnesses stay interchangeable, OpenAI just got free distribution. Those two strategies don't end in the same place.
Top developments
Anthropic — reversed course and extended Fable 5 access on all paid plans through July 19, plus kept Claude Code weekly rate limits 50% higher, hours after OpenAI temporarily lifted the 5-hour usage cap on Plus/Business/Pro and rolled out efficiency updates it says will stretch GPT-5.6 Sol further per query. Yesterday's exit-interview cycle became a same-week concession — frontier labs are now trading usage caps, not just price, for developer share. support.claude.com
Prime Intellect — shipped verifiers v1 wired into prime-rl, so a 100B-parameter reasoning model for 40-turn SWE-agent tasks (1,000 RL steps, sub-4-minute step time) now trains on 6 H200 nodes in under 2 days, packaged inside its Open Superintelligence Stack of training, inference, and compute. Custom-tuned SWE reasoners stopped being a hyperscaler-only game — a lab can now roll its own on rented compute over a weekend. x x
Notable discussions
The own-your-AI-stack turn — Satya Nadella's "Reverse Information Paradox" essay argues buyers give institutional learning back to the provider with every prompt, correction, and eval; David Sacks says most enterprise CTOs are trapped paying OpenAI/Anthropic because they can't wire the DoorDash-style token routing; cramforce reframes it as harness independence, not model independence; Aaron Levie says the corporate-IP layer is where value now compounds. Enterprise AI narrative is pivoting from best-model-wins to protect-your-learning-loop. x x x x
GPT-inside-Claude-Code goes official — Thomas Sottiaux (OpenAI's Codex lead) publicly told devs to point Claude Code at gpt-5.6-sol via CLIProxyAPI; hqmank and dr_cintas walked through the 5-minute setup; mweinbach admits most of his Claude Code usage has been GPT models for a while; cjzafir shipped a Codex-Orchestration plugin that lets Fable 5 plan and Sol high-effort execute inside one Codex session. OpenAI publicly endorsing its model running inside a competitor's harness closes the "which harness wins" question — harness is the runtime, model is the callable. x x x github.com
François Chollet — argues strong AI code gen has completely flipped its incidence: last year's weak models raised the floor for low-skill programmers and were useless to high-skill ones; today's strong models are most useful to high-skill programmers while low-skill ones underutilize or drown in them. Went from crutch to power tool. x
Andrew Ng — says 100% of his own coding tasks are now done by AI agents and gives it 3–6 months before prompting is gone, with self-improving loops as the next step. The Google Brain founder now sits fully inside the harness-runs-the-model camp. x
Tim Dettmers — reads OpenAI's Groq acquisition as targeting the MoE decode-layer bandwidth bottleneck: SRAM devices are cheap to manufacture and can serve distributed MoE effectively over strong interconnects, while HBM stays for attention decode. Hardware-inference co-development is where the next real cost win lives. x
John Crickett — "AI writes better code than I do" is a strange thing to say — AI was trained on published code and reflects the average of it, so the real statement is "my code is below average." Sharpest counter to the emerging just-let-it-code narrative. x
Pietro Schirano — ships a Coding Agent Index explorer with three surprises: Terra Max edges Fable 5 Max (77.4 vs 77.2) for ~76% less per task, Sol XHigh sits 1 point behind Max at ~26% lower API cost, and Luna Max beats Opus 4.8 Max for ~80% less per task. Fable's price/perf premium just evaporated at the API layer. api.magicpath.ai
Alex Finn — reads Anthropic's Fable 5 extension as conceding the AI race: nobody would pay Fable's API prices, so the alternative was Opus 4.8 vs GPT-5.6 Sol — the largest frontier-lab gap ever — and everyone would have switched overnight. Guarantees Opus 5 drops the moment Fable comes off subs, as the affordable Fable. x
Other news
Models & releases
NVIDIA/MIT SparDA — new transformer variant adds a Forecast projection that prefetches next-layer KV blocks, delivering 1.7x decode and +6.5pt long-reasoning gain at 0.41% param cost arxiv.org
Gemini 3.5 Pro leak — reportedly outperforms Fable 5 and GPT-5.6 in internal evals, targeting July 17 launch x
GPT-5.6 Sol Ultra math proof — Sam Altman resurfaces the 50-year-old Cycle Double Cover Conjecture result x
Grok 4.5 Pareto frontier — Sebastian Raschka charts it at the intelligence-per-cost frontier x
Anthropic Claude certifications — three official certifications launched x
Devtools & coding agents
Codex Chrome browser automation — Codex update enables web browsing via imported Chrome credentials x
Anthropic Loop Engineering course — free 45-minute AI fluency workshop from basics to production x
Anthropic Claude prompting workshop — 27-minute session on prompting techniques x
jakubkrehel skill pack — /better-ui, /better-typography, /better-colors skills for encoding judgment into agents github.com
Microsoft Claude Code rollout paper — 23-page PDF on Peer → Adopt → Loop → Ship spread across tens of thousands of engineers arxiv.org
Infrastructure & platforms
brianbellx GLM-5.2 compression — 423GB removed from a 753B-weight model bit-for-bit exact by keeping weights compressed in VRAM brianbell-x.github.io
Waterloo intern quantization — beats NVIDIA's official modelopt on runtime, aggressiveness, and benchmark score after three months of work x
Whisper large-v3-turbo on iOS — enables on-device speech-to-text in iOS apps x
Industry & policy
Anthropic hiring 2:1 over OpenAI — 891 vs 443 open roles across 19,026 pulled from 111 companies' own ATS; xAI actually shrinking x
Sam Altman on AI job creation — AI has been net-job-creating contrary to initial expectations x
S&P downgrades Oracle — flags OpenAI counterparty exposure as a key credit risk x
LangChain founder on context engineering — compaction + file systems + memory is the actual skill behind long-horizon agents x
Cursor / Cognition post-training — post-training work will be replicated industry-wide within months x
Continuing threads
Karpathy content wave (cont. from 07-12) — Karpathy's actual LOOPS.md workflow file from Anthropic surfaces publicly; MyWestLord/cyrilXBT reframe the Obsidian-as-IDE + Claude-as-programmer second-brain build (now 41K stars on kepano's skill pack); Karpathy tells builders to use AI as a thinking partner, not a code writer x x x
Claude Code in-app browser (cont. from 07-11 Top dev) — desktop browser rollout being amplified across the developer feed x