AI Coding Model Reshuffle | AI Field Notes #54

AnalysisThe highest-scoring coding model on the public leaderboards is the one developers cannot reliably touch. Anthropic's Fable 5, which posts 80.3% on SWE-bench Pro (a test of whether a model can fix real software bugs), spent most of June offline after a US Commerce Department export directive on June 12 cut access for foreign nationals, and the company's free trial expired on June 23. It returns priced at $10 per million input tokens and $50 per million output, double Claude Opus 4.8. Anthropic told customers the models would return 'in the coming days,' a phrase now ten days old.

AnalysisNvidia just got designed out of the largest AI infrastructure plan of the year. China committed roughly $295 billion over five years, about $59 billion annually, to state-directed AI data centers built on domestic accelerators rather than imported silicon, according to reporting on June 22. The timing was pointed: it landed ten days after Washington's export directive against Anthropic and during G7 talks where President Trump called chip-export discussions 'going fine.' One Chinese AI chief said his company would match Fable 5-class capability before the end of the first quarter of 2027.

AnalysisForty-two state attorneys general have opened a coordinated investigation into OpenAI, with New York's office already issuing subpoenas. The questions cover advertising claims, ChatGPT's tendency toward flattery (the 'sycophancy' problem, where a model tells users what they want to hear), how the company handles personal and health data, and protections for minors and older users. The timing matters because OpenAI is steering toward a public offering later in 2026, and a multi-state legal action is the kind of disclosure that gives bankers and prospective shareholders pause.

AnalysisGoogle moved Gemini 3.5 Pro into general availability this week, giving developers a frontier model with a 2 million token context window, the largest in production, meaning it can hold roughly a mid-size code base or a stack of long documents in a single request. API pricing runs about $15 per million input tokens and $60 per million output. The most capable setting, a slow 'Deep Think' reasoning mode, sits behind a $250-a-month Ultra subscription. With Fable 5 hobbled by export rules, Google's timing puts a stable, fully available option in front of teams shopping for a new default.

AnalysisGoogle shipped Android 17 with its Gemini models wired into the operating system itself rather than bolted on as a separate app. The headline piece, Gemini Omni, handles text, images, audio, and video together, with on-device translation and music generation running locally instead of in the cloud. The part developers care about: third-party apps can reach these abilities through new Android AI interfaces, so a small team can add live translation or multimodal search without training or hosting a model. It also deepens the dependency, since the smartest path on Android now runs through Google's stack.

AnalysisDevelopers building on Amazon's cloud can now call xAI's Grok 4.3 directly through Bedrock (AWS's menu of hosted AI models), under the model ID xai.grok-4.3. Pricing comes in low for a frontier model, around $1.25 per million input tokens and $2.50 per million output, with a 1 million token context window and adjustable reasoning depth. xAI claims the lowest hallucination rate among leading models, a claim worth testing rather than trusting. The real shift is distribution: Grok now sits in the same console as Anthropic and Meta models, one API call away from any AWS shop.

AnalysisNearly every working developer now uses an AI coding tool, 97% by a Black Duck survey (Black Duck is a software-security firm), with GitHub Copilot at 83% and Anthropic's Claude Code at 63%, striking for a product not yet a year old. The gap is governance: only about a third of organizations have rules for how that generated code gets reviewed, licensed, or secured before it ships. Adoption ran ahead of policy, which is how a tool becomes load-bearing before anyone decides who is accountable when it is wrong.

AnalysisDeepSeek previewed V4, an open model with 1.6 trillion parameters in a mixture-of-experts design (which activates only a slice of the model per request to cut cost), and the notable detail is the silicon. The Chinese lab says it trained V4 on Huawei's Ascend accelerators rather than Nvidia hardware, the exact dependency US export controls were built to exploit. Early reads call it the strongest open model available, though still short of the best closed US systems. The weights are downloadable, and the training story is the warning shot.

AnalysisOpenRouter, a service that routes API calls across many AI models, released Fusion, which combines several models' answers into one on the server side and hands it back like a single model. In one published test, a budget panel of Gemini 3 Flash, Kimi K2.6, and DeepSeek V4 Pro scored 64.7% on a reasoning benchmark, within a point of Fable 5's 65.3%, at about half the cost. The trick is old, ask several models and merge the best, but packaging it behind one API call makes it something a small team can actually use.