AI model
OpenAI's GPT-5.6 model family, released in June in Sol, Terra and Luna tiers from premium to low-cost, high-volume use; succeeded by GPT-6.
OpenAI roughly matched Anthropic in September on spending by about 120,000 shared customers, up from a quarter at the start of the year, OpenRouter data shows. Both labs are now cutting small-model prices before planned IPOs, so share costs each of them revenue per request.
Reality
USA TODAY Co. sued OpenAI for more than $250 million and an injunction over training ChatGPT on work from 19 publications. The case tests whether a publisher can force payment and limits on the training data that teams building on OpenAI's models depend on.
Perspective Coverage
5 publishersReality
OpenAI says GPT-6 Sol, for demanding professional work, and GPT-6 Luna, for high-volume jobs, cost up to 50% less than comparable GPT-5.6 options. The split is by workload, so each job has to be tested on its own before a team counts on the full discount.
Reality
Adversa AI got GitHub Copilot CLI to decrypt and run a hidden instruction, and Microsoft's mai-code-1.1-flash fell for it 50 percent of the time. GitHub declined to treat it as a product vulnerability, so teams running autopilot have to check what the agent executes themselves.
Reality
GitHub is rolling GPT-6.1 Sol into Copilot, citing early tests in which it used fewer tokens and steps than GPT-6 and GPT-5.6, without figures. Teams will have to measure any saving on their own tasks, at a rate the announcement did not quote.
Reality
One OpenAI Codex prompt spawned 826 child agents and burned about $78,000 in credits, according to the user's own reconstruction. Nearly all the counted tokens trace to an alpha client build, and only OpenAI's servers can turn them into dollars.
Reality
Microsoft's new Copilot app bills coding and agents by use, while fewer than 7% of over 450 million commercial 365 seats pay for Copilot. Usage billing lets Microsoft charge its existing Copilot customers more whether or not the unlicensed majority ever signs up.
Reality
Simon Willison's teardown finds a cloud container whose outbound domain list appears open by default, plus a desktop build that runs programs on the employee's machine. Those are two provisioning reviews with different threat models.
Reality
A dev.to guide to GPT-5.6 pricing shows batch halving both token rates and a cheap-first cascade saving money until 9 in 10 calls escalate. Batch is opt-in and caching fails silently on short prefixes, so the default request often pays list price.
Reality
Tuesday's model releases from Anthropic and OpenAI came with audit numbers that move in both directions at once. For anyone granting an agent tokens or tooling, a version bump means re-running the injection tests.
Reality
OpenAI's guide prices reused input tokens at a discount of up to 90 percent. It also says cached key-value states sit on individual machines, so an unchanged prefix can still miss the cache when routing overflows.
Reality
Tencent's Hunyuan Speech team put a swappable agent model behind a conversation model that re-decides every second whether to talk, and reports best-in-test timing on Full-Duplex-Bench v3 with a task-accuracy shortfall it describes only as slight.
Reality
XDA Developers handed Claude Code, Codex and Google Antigravity the same demanding site brief with no follow-ups. All three shipped every section that was asked for. The distance between them showed up in typography, spacing and mobile mockups.
Reality
California's transparency law makes a developer's own safety policy enforceable against it, and the Midas Project built its case entirely from OpenAI's Frontier Governance Framework and the system cards published after it.
Reality
Commerce and the White House put three named frontier models from the two leading US labs under access restrictions in June 2026. Enterprise access to those models now runs through a government approval step.
Reality
Jensen Huang says OpenAI's Astra reached AGI, and the tape answered by putting 15% into CoreWeave and taking 2% out of Nvidia, which says rather more about a $108 billion revenue guide than about machine cognition.
Reality
The same guide that tells you to run evals on every change now carries a deprecation notice with two dates. For anyone whose deploy gate creates eval runs, the earlier date is the one that bites.
Reality
OpenAI's create response reference is explicit that a chained request does not inherit earlier instructions, and because the field is optional, the turn that loses your JSON contract or your safety text still succeeds.
Reality
SemiAnalysis counts roughly $350m of state money behind five competing Korean foundation-model consortiums, set against a $919bn infrastructure figure. The gap is where the investment case sits, and it points at hardware.
Reality
Anthropic's revenue keeps climbing, on the FT's sourcing, while the card-billing index Ramp builds from 70,000 companies shows July's Anthropic spend going to cheaper Claude models instead of the flagship.
Reality
Product · August 14, 2026 · Confirmed2 publishers