AI News — June 26, 2026

Model releases

• [2026-06-24] OpenAI — Jalapeño, OpenAI’s first custom chip (which it brands an “Intelligence Processor”), co-designed with Broadcom: an ASIC architected specifically for LLM inference, taken from initial design to manufacturing tape-out in nine months — described as possibly the fastest ASIC development cycle in high-performance semiconductors. OpenAI says early testing shows performance-per-watt substantially better than current state-of-the-art (a full technical report is promised “in the coming months”), with initial deployment targeted for end of 2026, ramping through 2027 into 2028. It matters because OpenAI moving to serve ChatGPT on silicon it designed itself is a direct bid to cut inference cost and reduce dependence on Nvidia GPUs — vertical integration of the inference stack. (official, CNBC, TechCrunch)

  For anyone buying OpenAI tokens the payoff is indirect and a year out — inference OpenAI runs on a chip it designed is cheaper to serve, which is eventually what funds lower prices or higher rate limits — while the nearer signal is the nine-month design-to-tape-out itself: evidence that custom inference silicon has gotten fast enough to build that renting Nvidia GPUs is no longer the only way a lab will plan to serve a model.

Coding agents

• [2026-06-24] Anthropic — Claude Code v2.1.191 (and v2.1.190) added /rewind support for resuming a conversation from before /clear was run, made MCP capability-discovery and OAuth requests retry transient network errors with short backoff, cut CPU usage during streaming by ~37% (by coalescing text updates to 100ms) and reduced long-session memory growth, and fixed background agents resurrecting after being stopped (stopping from the tasks panel is now permanent). The /rewind-past-/clear recovery and the MCP retry hardening are the substantive parts — they close two common ways a long session loses state or stalls on a flaky connection. (official)

  Most of this release is reliability plumbing rather than new capability, and the MCP retry hardening is the quietly important piece — capability discovery and OAuth now survive a dropped packet, so a flaky VPN or corporate proxy stops turning a transient blip into a dead session.

  For Software Developers: If you’ve ever run /clear to shed a bloated context and only then realized you still needed a file path or decision from the conversation you just wiped, /rewind can now step back across that boundary to recover it — the earlier turns are reachable again instead of gone for good.

• [2026-06-24] GitHub — Copilot Free and Student plans switched to Copilot auto model selection as the default and only model-selection experience, removing manual model choice; auto routes each task to a model across families (subject to plan limits). GitHub also retired the “(Preview)” label from Microsoft-released models, saying the router’s continuous behind-the-scenes improvements make the label unnecessary. It matters because on the free tiers you no longer pin a model — task and prompt design now have to assume a router’s pick rather than a known model. (official)

  On the Free and Student tiers, prompt design can no longer assume a known model: a prompt shaped around one model’s habits now has to hold up when the router picks a different family for the same task, and with the “(Preview)” labels gone there’s also no longer a cue telling you a given Microsoft model is still experimental.

Watch list

• Alibaba / Qwen distillation fallout — legislative action is now moving off Anthropic’s complaint that Alibaba-linked operators ran 28.8 million exchanges through ~25,000 fraudulent accounts: Senators Bill Hagerty (R-TN) and Andy Kim (D-NJ) are preparing an amendment to must-pass defense legislation that would blacklist or sanction any Chinese firm found improperly accessing US AI-model output to train competitors. Watch for whether the amendment is formally introduced, the administration’s response to Anthropic’s “coordinated government–industry” ask, and whether Alibaba (still silent) responds. (CNBC, prior coverage)

  Because the amendment would ride on must-pass defense legislation, it is far harder to strip than a standalone bill — so the milestone that actually moves this is Hagerty and Kim formally filing it, after which the fight shifts from whether to sanction to which Chinese firms the language reaches.

• Fable 5 / Mythos 5 restoration — both still suspended under the June 12 export-control directive, now 14 days on with no restoration date; Anthropic’s notice still commits only to restoring access “as soon as possible,” and usage-credit metering has been live since June 23. The updated consumer terms enumerating identity-verification data take effect July 8, which analysts read as the likely mechanism for a US-users-first return. Watch for a date and whether access comes back gated by identity or nationality. (official notice, terms update, prior coverage)

  July 8 is the date to circle: if access returns gated by verified identity, that is the same attributable-access machinery yesterday’s distillation story points to — arriving not as a one-time unlock for Fable 5 but as a standing condition of frontier access.

• Gemini 3.5 Pro (late June) — still a limited Vertex AI enterprise preview with GA targeted for June; with month-end days away there’s still no model card, public benchmarks, or pricing for the promised 2M-token context and Deep Think mode. The open question is whether Google holds the June date or lets it slide into Q3. (prior coverage)

  Month-end is now the real deadline — if June closes with the preview still enterprise-gated and no model card, the story stops being “running late” and becomes whether Google has quietly reset the GA target to Q3 without saying so.

• xAI Grok V9-Medium — the ~1.5-trillion-parameter, Cursor-trained coding model is reportedly live in the consumer Grok product on X and SuperGrok but still has no API string or model card on x.ai, so nothing testable has shipped for anyone building on it. (prior coverage) (unconfirmed)

  A model can be “live” for consumers on X and SuperGrok while staying entirely unbuildable — with no API string or model card on x.ai, every claim about V9-Medium’s coding ability is, for now, unverifiable by anyone outside xAI.

• 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)

  Beyond the feature list, the formal deprecation policy is the part with lasting reach: once it lands, MCP finally has a defined support window, which would make 2026-07-28 less a feature drop than the first version you can build against without budgeting for breaking churn every few months.