OpenAI ships GPT-5.6 behind a government gate, the same day Washington un-blocks Mythos 5

OpenAI shipped a frontier model yesterday and then said almost no one can use it. GPT-5.6 arrived as a three-model family (Sol the flagship, Terra a balanced mid-tier, Luna the cheap high-volume tier), but as a limited preview restricted to a small group of trusted partners in Codex and the API, explicitly “at the request of the U.S. government.” Reporting relayed by commentators put the initial pool at roughly 20 government-approved companies, with general availability promised “in the coming weeks.” Sam Altman said OpenAI had planned a broader launch and pulled it back to a preview after the government request. The model itself is the strongest OpenAI has shipped: its preview post claims Sol sets a new state of the art on Terminal-Bench 2.1, with Sol Ultra reaching 91.9%, and posters noted Terra is the first flash-sized model above 80% on that benchmark. Pricing puts Sol at $5/$30 per million input/output tokens, Terra at $2.50/$15, and Luna at $1/$6, which lands Sol above Claude Opus 4.8 on output cost ($5/$25) but well under Mythos 5 ($10/$50), with Luna roughly at GLM-5.2’s blended ~$2. OpenAI was careful to say Sol does not cross the Cyber Critical threshold under its Preparedness Framework: in Chromium and Firefox evaluations it surfaced bugs and exploitation primitives but did not autonomously produce a full-chain exploit. Two runtime knobs shipped alongside, “max reasoning” for a longer deliberation budget and “ultra mode” that spins up subagents, the kind of harness-level orchestration agent teams had been building themselves.

The release reads differently next to what Anthropic posted the same day. Anthropic said the U.S. government notified it that Mythos 5, its strongest cybersecurity model, can be redeployed to a set of U.S. organizations that operate and defend critical infrastructure, two weeks after the June 12 controls that pulled Mythos 5 and Fable 5 and about a week after Anthropic floated a restoration proposal to Commerce. Fable 5 is still waiting on general availability. “The US lifts its block on Mythos 5” ran on Hacker News’ front page and across Techmeme. Two of the three frontier labs spent the same day routing model access through Washington, one launching gated and one getting partly un-gated, both at government direction. Trusted-partner-first deployment is starting to look like the default path for a frontier release rather than the exception.

While access tightens at the top, the economics underneath keep pushing the other way. Coinbase cut its AI spend nearly in half while token usage kept climbing, and Brian Armstrong laid out how: default engineers to open-weight models (GLM 5.2 and Kimi 2.7) through an internal gateway rather than capping usage, since 91% of employees never hit their caps anyway; route by task, sending a frontier model the planning and a cheaper one the execution; and fix caching, which took their LibreChat cache-hit rate from 5% to 60%. Engineers can still pick any model; the defaults carry the savings. Gergely Orosz, who surfaced the writeup, asked whether this is the start of a trend, and given that the week’s other running story was open models reaching parity on coding benchmarks, the answer looks like yes.

The labor picture got fresh numbers from Anthropic’s June Economic Index, the first to survey Claude users directly alongside hourly usage sampling. Over a third of respondents expect AI to handle most or nearly all of their work tasks within a year. Fewer than 10% think they’ll lose their own job in that window, but more than a third put the odds of a junior colleague losing theirs above 60%, and the people delegating the most work to AI were the most optimistic about their own pay and security. The index now also tracks artifacts, the concrete output of a session, a more meaningful unit than raw message counts.

Tooling caught up to how people are actually running agents. GitHub Desktop 3.6 added Git worktree support to the GUI, which the team framed around coding agents that “spin up worktrees to run isolated, parallel sessions.” Copilot in Desktop now runs on the Copilot SDK, with commit-message generation that reads your copilot-instructions.md and AGENTS.md and honors repo commit-metadata rules, AI-assisted merge-conflict resolution, a model picker, and BYOK for local or third-party models.

Two practitioner posts pointed at where long-context work is heading. A developer published ContextForge (arXiv 2606.26105), which treats the context window as a working set rebuilt minimally at each step instead of a transcript carried forward, adds a queryable “LLM Wiki” memory layer, and reports steadier answers and token usage across 180- and 500-day synthetic runs measured against RecallBench’s temporal-reasoning and contradiction-resolution tasks; it runs locally on a SQLite store with no vector DB required. And an SRE fed 850k tokens of pod logs, Prometheus exports, the Slack incident channel, and Jira comments into MiniMax M3 and got the root cause of a cascading 502 incident in about 90 seconds: a 6-hour cron job triggering HPA scale-downs that collided with a 15-second graceful-shutdown window against 30-to-45-second requests. His team had taken 14 hours across two days to reach the same chain. He flagged it wasn’t a blind test, since he already knew the answer, but the cross-referencing speed across mixed-signal data is the part worth watching.

What to watch: whether OpenAI and Anthropic hit their “coming weeks” timelines for general access or whether trusted-partner gating sticks, and whether Coinbase’s open-weight-by-default playbook shows up in other engineering orgs’ spend next quarter.