GPT-5.6 Goes Public | AI Field Notes #64

AnalysisEveryone can use GPT-5.6 starting today, two weeks after OpenAI first showed it to roughly twenty vetted partners and no one else. The wait was not technical. The Commerce Department's Center for AI Standards and Innovation ran its own tests before the Trump administration signed off on a broad release. Three models arrive: Sol at $5 and $30 per million input and output tokens, Terra at $2.50 and $15, and Luna at $1 and $6. Terra matches last year's GPT-5.5 at half the price, which is the part a team watching its inference bill will feel first.

AnalysisGrok 4.5 resolves a coding task in an average of 15,954 output tokens where Claude Opus 4.8 burns 67,020, a gap of more than four to one that lands on the bill for anyone running coding agents at volume. xAI shipped it July 8, built on a 1.5-trillion-parameter mixture-of-experts foundation (a design that fires only part of the model per request to cut cost) and trained with Cursor on real coding sessions. It runs near 80 tokens a second at $2 and $6 per million, and tops Harvey's legal-agent benchmark. It beats Opus on two of the four tests xAI chose to publish and loses the other two.

AnalysisMost agents still fail on reliability rather than raw intelligence, and Bespoke Labs just raised $40 million to attack exactly that. The company builds training environments that mimic a real workplace: a live codebase, an inbox, a Slack history, the messy context an agent has to act inside. Post-training runs against those simulations so a model learns to behave, not just answer. The round, led by Wing VC with Mayfield and engineers from Anthropic, follows an earlier $8.25 million backed by Google DeepMind's Jeff Dean. The bet is that the next capability jump comes from better practice grounds, not bigger models.

AnalysisJosh Kushner's Thrive Holdings is raising about $2 billion from SoftBank, Altimeter, and D1 Capital to do something more direct than sell software to accountants: buy the accounting firms and automate the work from inside. OpenAI holds a stake and has lent researchers to the effort, including its head of applied research, Boris Power. One portfolio company, Current, has already acquired 48 accounting firms, and Thrive staff say they built a tax-return-processing agent on OpenAI's Codex that Current now runs. The pitch to investors is margin: strip repetitive human work out of a business you own outright.

AnalysisBuilding a phone agent that sounds human used to mean stitching together speech models, telephony, and a pile of glue code. xAI now sells the whole stack as one no-code product. Its Voice Agent Builder, in beta since July 1, bundles telephony, retrieval, tools, guardrails, and observability, and this week added 21 multilingual voices with cloning. Pricing runs 5 cents a minute of audio plus a cent for phone numbers. Grok Voice scores 67.3% on the tau-voice benchmark against 43.8% for Gemini 3.1 Flash Live, so the quality claim has a number behind it.

AnalysisTo keep using Claude's best models, a freelancer on a Pro plan may now have to hand over a government photo ID, a live selfie, and a scan of their facial geometry, run through a third-party vendor called Persona. Anthropic's updated policy took effect July 8 and covers Free, Pro, and Max accounts; the API, Team, and Enterprise tiers are exempt. The trigger was a June 12 export-control order requiring Anthropic to keep foreign nationals off its most capable systems. With no way to check nationality, it built face verification instead, and now concedes the data may be considered biometric in some places.

AnalysisMicrosoft wired its Dataverse business data directly into Claude, Cursor, and GitHub Copilot this week, and the interesting part is the governance layer around it. The plugin speaks the Model Context Protocol (MCP), the shared standard that lets an agent reach a tool or data source, so a developer can query live company records from inside their coding assistant. Administrators keep control of what each agent can touch, with auditing on every call. That combination, live access with a permission boundary, is what has kept regulated teams from pointing agents at production data. Microsoft is betting the boundary is the product.

AnalysisA Chinese open-weight model quietly showed up on GitHub Copilot's menu this week: Moonshot AI's Kimi K2.7 Code now sits alongside the usual OpenAI and Anthropic options for millions of developers. Microsoft also moved Copilot Vision to general availability, so the assistant can read screenshots and diagrams, and shipped CLI version 1.0.69 with a plugin dashboard and hot-reloading extensions. None of it is a headline release. Taken together it shows the shape of things: the model roster is becoming a commodity shelf, and a well-priced open model from Beijing gets stocked next to the rest.

AnalysisTalking to ChatGPT stopped feeling like trading walkie-talkie turns today. OpenAI began rolling out GPT-Live, a pair of voice models (GPT-Live-1 and a smaller mini) built on a full-duplex design, meaning the system can listen and speak at once. It backchannels with an 'mhmm' while you think, jumps in when it has something, or holds quiet, and it reaches CarPlay. The mini is the default for free users; paid tiers get the larger model. An API is promised soon. The consumer bar for what a spoken conversation with software should feel like just moved, and rivals will be measured against it.

AnalysisThe reason GPT-5.6 sat behind a twenty-partner wall for two weeks is a government approval process that barely existed a year ago. Before a broad release, OpenAI's model went through the Commerce Department's Center for AI Standards and Innovation, and a wider framework lands August 1: classified benchmarking run with the NSA and CISA, plus a voluntary pre-release review that decides which systems count as covered frontier models. Google's Gemini 3.5 Pro is caught in the same queue. A compliance layer is forming between finishing a frontier model and letting the public use it, and it is being built in real time.

AnalysisTwo days before the consumer voice upgrade, OpenAI shipped the developer half: GPT-Realtime-2.1 and a mini variant, voice models aimed at the API for anyone building their own agents. The pitch is lower latency and added reasoning, the two things that make a spoken agent feel like it is following the conversation rather than transcribing it. It landed the same week xAI put voices behind its no-code builder, which is the tell. Voice agents are moving from a research demo to a stocked shelf of parts, priced per minute and per token, that a small team can assemble without a speech PhD.