AI News — June 27, 2026

Model releases

• [2026-06-26] OpenAI — GPT-5.6 previewed as a three-tier family: Sol (flagship, for the hardest coding and security-research problems), Terra (balanced, for high-volume business work — roughly 2× cheaper than GPT-5.5 with competitive performance), and Luna (fastest, lowest-cost). Pricing per 1M tokens: Sol $5 in / $30 out, Terra $2.50 / $15, Luna $1 / $6. OpenAI classified all three at “High” risk for cyber and biological/chemical capability (internal capture-the-flag cyber scores of 96.7% / 91.84% / 85.19%), and — at the US government’s request, after previewing the models’ capabilities ahead of launch — is starting with a limited preview to ~20 trusted partners whose participation was shared with the government, with general availability “in the coming weeks.” OpenAI tied the staggered rollout to the June 2 executive order on frontier-model benchmarking and said such restrictions “shouldn’t be the norm.” It matters because OpenAI is now releasing a flagship behind a government-coordinated gate rather than to the public — the same access-by-permission pattern Anthropic’s Mythos/Fable saga has been tracing all month. (official, TechCrunch, VentureBeat)

  Even the lowest-cost tier, Luna, lands at “High” cyber risk and inside the same ~20-partner gate, so this isn’t a flagship-only restriction — the whole family is uncallable for anyone building on it until the “coming weeks” GA. The one number you can act on today is the price card: Terra at $2.50/$15 undercutting GPT-5.5 by roughly half is OpenAI signalling where it expects mid-tier inference to settle once the gate lifts.

• [2026-06-26] Anthropic — the US government partially lifted the June 12 suspension of Mythos 5. Commerce Secretary Howard Lutnick notified Anthropic (in a letter to chief compute officer Tom Brown) that Mythos 5, its strongest cybersecurity model, may be redeployed to a curated set of roughly 100+ US organizations — government agencies and private companies that operate and defend critical infrastructure — for defensive cyber use. It is a marked softening of the directive imposed two weeks ago after researchers showed the models’ guardrails could be bypassed. Notably, the letter does not mention Fable 5, which remains suspended (now day 15); people close to the talks say a Fable release is being worked toward but the timeline is unclear. It matters because the restoration comes back gated by org and use-case rather than as a general unlock — and arrived hours after OpenAI’s government-coordinated GPT-5.6 limit, so on the same day both leading US labs shipped frontier capability only to vetted lists. (Bloomberg, 9to5Mac)

  Mythos 5 — the more capable cyber model — returning before consumer-grade Fable 5 inverts the usual “ship the safe one first”: routing the stronger model to a vetted defender list is easier to justify than a public unlock. For the roughly 100 named organizations it means their strongest defensive-cyber tool is back, scoped to defensive use; for everyone outside that list, nothing changes yet.

Coding agents

• [2026-06-26] Anthropic — Claude Code v2.1.193–v2.1.195 shipped. The substantive additions are in v2.1.193: a new claude_code.assistant_response OpenTelemetry log event that captures the model’s response text (so audit/telemetry pipelines can log what the agent actually said), an autoMode.classifyAllShell setting that routes every Bash/PowerShell command through the auto-mode classifier, auto-mode denial reasons now surfaced in the transcript, denial toast, and /permissions recent denials, and live file-path autocomplete in bash (!) mode. v2.1.195 is mostly fixes — hook matchers with hyphenated identifiers (e.g. code-reviewer, mcp__brave-search) now exact-match instead of substring-matching, plus background-job and macOS voice-dictation repairs. The OTel response-text event and the classify-all-shell toggle are the parts worth noting — both tighten observability and policy enforcement around what the agent does, rather than adding new capability. (official)

  Hyphenated hook matchers behave differently after v2.1.195: a matcher like code-reviewer or mcp__brave-search that was silently catching every identifier containing that substring now fires only on an exact match, so any hook config using hyphenated names is worth re-checking after the upgrade.

  For Security Engineers: Flip on autoMode.classifyAllShell and every Bash/PowerShell command an agent issues hits the auto-mode classifier before it runs — closing the gap where only some commands were checked — and the new claude_code.assistant_response OTel event records the model’s actual response text, so an auto-mode deployment becomes auditable end-to-end (the policy check plus a logged record of what the agent said) rather than spot-checked.

Watch list

• Fable 5 general release — with Mythos 5 cleared for critical-infrastructure defenders on June 26, Fable 5 (the consumer-grade sibling) is the piece still dark, 15 days into the blackout and absent from the government’s clearance letter; people close to the talks say a release is being worked toward with no committed date. The July 8 consumer-terms update enumerating identity-verification data remains the most-cited mechanism for a US-users-first return. (Bloomberg, prior coverage) Fable 5 is the last frontier piece still dark, and the detail to watch is less the date than the terms: a return gated by verified identity (the July 8 consumer-terms path) versus a plain public unlock sets the template for how consumer frontier access resumes after a blackout.

• GPT-5.6 general availability — OpenAI says Sol, Terra, and Luna move from the ~20-partner limited preview to general release “in the coming weeks”; watch for the GA date, whether the government-coordination gate persists at GA, and published benchmarks beyond the internal CTF cyber scores. (official) Whether the government-coordination gate survives to general availability is the real question: if GA still ships through a vetted-partner channel, “limited preview” stops being a launch phase and starts looking like the steady state for a model family classified “High” on cyber.

• Gemini 3.5 Pro (GA) — still a limited Vertex AI enterprise preview with no model card, public benchmarks, or pricing for the promised 2M-token context and Deep Think mode; Sundar Pichai’s “give us until next month” from Google I/O now lands at month-end. Watch for whether June closes with a rate card or the GA target quietly slides into Q3. (prior coverage) June closes in days with still no model card, benchmarks, or pricing, so this comes down to a binary: a rate card this week, or a GA target that has quietly slid to Q3 — and a third straight month of enterprise-only preview would put Pichai’s “next month” two deadlines behind.

• Alibaba / Qwen distillation amendment — the Hagerty–Kim amendment to must-pass defense legislation (which would blacklist or sanction Chinese firms found improperly accessing US AI-model output) was reported as expected “as soon as Wednesday” (June 24); watch for whether it has been formally filed and Alibaba’s still-absent response. (prior coverage) The amendment was floated for June 24 and that date has now passed without a confirmed filing, so the thing to track is whether it has actually been entered into the defense bill’s text — riding must-pass legislation is what makes it hard to strip, but only once it’s formally filed does the fight move from “whether” to “which firms.”

• MCP spec finalization (July 28) — the 2026-07-28 release candidate (stateless core, Extensions framework, Tasks, MCP Apps, authorization hardening, formal deprecation policy) is in its validation window; watch for the rest of the SDK support landing before the cutover. (official, prior coverage)

  A release candidate in its validation window means the wire format is effectively frozen, so this is the last stretch where a breaking problem found in Tasks, MCP Apps, or the authorization hardening can still change the spec before July 28 — server and client maintainers testing against the RC now have more leverage to surface issues than they will after the cutover.