AI News — July 8, 2026

Model releases

• [2026-07-06] OpenAI — gpt-realtime-2.1 and gpt-realtime-2.1-mini shipped on the Realtime API. The update brings reasoning to realtime voice — with configurable reasoning effort and a spoken preamble so the model can keep talking while it works through a multi-step request — and improves alphanumeric recognition, silence/noise handling, and interruption behavior; OpenAI says it cut p95 latency by at least 25% across its Realtime voice models through improved caching. The mini tier is a distilled reasoning model aimed at faster, lower-cost voice agents. It matters because it pushes the Realtime API past a transcription-and-TTS pipe toward voice agents that can reason and use tools mid-conversation, and pairs that with the latency and cost cuts that make always-listening agents viable in production. (official, docs)

  Adding reasoning to a live voice call creates a latency problem the text API never had — a caller won’t sit through multi-second silence — so the spoken-preamble behavior, where the model keeps talking while it works through the request, is doing as much for the experience as the reasoning itself, and it’s the first thing worth stress-testing on real calls.

  For Software Developers: If your voice agent is a transcribe → model → TTS pipeline today, gpt-realtime-2.1 lets you collapse it into one Realtime session that reasons and calls tools mid-turn — a support line that looks up an order mid-sentence and fills the wait with a preamble instead of dead air — and the distilled mini tier gives you a cheaper model to route the high-volume, low-complexity calls to.

AI cost tracking & telemetry

• [2026-07-07] US companies are shifting AI spend toward cheaper Chinese open-weight models as frontier prices climb. Per CNBC, the share of tokens US companies run on Chinese models via OpenRouter has stayed above 30% every week since February 8 and spiked as high as 46%, against an ~11% average over the prior 12 months. The driver is cost: open-source Chinese models run 60–90% cheaper than the leading OpenAI and Anthropic models, and Z.ai’s GLM 5.2 landed within a percentage point of Claude Opus 4.8 on one closely watched agentic benchmark at roughly a fifth of the cost — posting the fastest adoption of any model Vercel tracked in 2026 (daily token volume ~27×, customers ~80× in its first full week). Named migrations: AI startup Lindy moved 100% of its traffic from Claude to DeepSeek in June, while Shopify and Airbnb have touted Alibaba’s Qwen 3. It matters because it puts a market number on the cost thread this week has tracked — Fable 5’s $10/$50 rate, workspace-agent metering — showing frontier-tier price hikes translating directly into US buyers substituting toward open-weight Chinese models rather than absorbing the increase. (CNBC)

  For teams doing the substituting, the enabler underneath these numbers is OpenRouter-style routing itself — CNBC can measure the shift precisely because swapping a Chinese open-weight model in behind the same API surface is now a config change, not a migration — which cuts both ways: a 30–46% share that moved this fast can swing back just as fast if a US frontier vendor drops prices or one of these models regresses under real workloads.

Watch list

• Fable 5 residuals — the pricing question closed July 7 ($10/$50 per million tokens via metered usage credits). The open residuals are behavioral: whether the government-coordinated cybersecurity classifier over-blocks legitimate red-team and security work now that the model is in steady-state paid use, and whether standard Enterprise seats — which carry no included Fable 5 allowance — can enable usage credits at all. (official, prior coverage)

  Pricing is settled, so what’s left to watch is behavioral: whether the cyber classifier over-blocks legitimate red-team work now that the model is in steady-state paid use, and whether Enterprise seats can turn on usage credits at all — the second answer decides whether Fable 5 is even reachable for those users, regardless of the published rate.

• GPT-5.6 general availability — Sol, Terra, and Luna remain a ~20-partner, government-approved limited preview; preview pricing is already public (Sol $5/$30, Terra $2.50/$15, Luna $1/$6 per million input/output tokens), and OpenAI still says broad GA is “in the coming weeks.” The gate is now explicit: Trump’s AI-security executive order directs the administration to stand up a classified process by August to assess models’ cyber capabilities and designate “covered frontier models,” so a clean public launch likely waits on that framework landing. Watch the GA date and the first third-party benchmarks beyond OpenAI’s internal CTF cyber scores. (preview, prior coverage)

  Today’s bullet names the actual gate: the executive order ties a clean public launch to a classified cyber-assessment process that isn’t due to stand up until August, so “coming weeks” realistically reads as “after that framework lands” — which makes the August deadline, not OpenAI’s own roadmap, the date to track.

• 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. Reporting now points to a mid-July target (around July 17) tied to a full architectural rebuild — scrapping the 2.5 Pro base for a fresh pre-training cycle — amid senior-researcher departures from DeepMind, though Google has confirmed none of it. Watch for whether even a model card lands this month or GA slips into Q3. (unconfirmed) (prior coverage)

  Reporting now pins the delay on a full re-pretraining that scrapped the 2.5 Pro base rather than a routine slip — unconfirmed, but if it holds up it explains the total public silence, and it means even a bare model card landing this month would be the first real sign the rebuild actually converged.

• 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 with the Ruby/TypeScript/Python SDKs updating against it. Watch for the remaining SDK support landing — and any breaking issues in Tasks, MCP Apps, or the new MCP-specific HTTP headers — before the cutover. (official, prior coverage)

  With the RC in its validation window, the piece most likely to move the July 28 date is the SDK side: if Ruby, TypeScript, and Python don’t all land clean support against the stateless core and new HTTP headers in time, it’s the cutover that slips, so the SDK pull requests are a better tell than the spec text.