The AI Wire ●

Xiaomi published the weights for MiMo-V2.6-Pro and MiMo-V2.6-Flash, its newest open-weight multimodal models, alongside the reinforcement-learning machinery used to build them — a materially broader disclosure than a typical API launch. Both models accept text, images, video and audio with a claimed 1-million-token context window: Pro is a sparse mixture-of-experts with 1.02 trillion total parameters (42 billion activated per token), Flash lists 309 billion total with 15 billion activated.

The release package includes a technical report, deployment instructions and RL system details, and Xiaomi says it is also open-sourcing the training environments and RL code needed to examine and reproduce the work — though serving models this size still requires distributed multi-GPU inference. Xiaomi is also offering a hosted MiMo-V2.6-Pro-UltraSpeed tier it claims generates output up to 20x faster than standard Pro service at the same quality, and the models are listed for hosted access on OpenRouter.

Fresh third-party validation (Sept 22): Artificial Analysis scores MiMo-V2.6-Pro at 46 on its Intelligence Index v4.3.2 — the highest open-weights result on that leaderboard, ahead of Z.AI's GLM-5.3 (45) and Kimi K3 (44). Xiaomi claims a 46.32 and calls the model 'the strongest open-source model to date,' keeping the V2.5 series' API pricing — which it says pushes the intelligence-versus-cost Pareto frontier outward. Among proprietary systems, the same index lists Grok 4.7 at an identical 46, putting the open model level with xAI's flagship on this benchmark.

StepFun's new flagship is a sparse MoE (600B total, 27B active) with 1M-token context and image/video input, priced at $1 per million input tokens — undercutting frontier rivals — and aimed at long-horizon agentic work. It is live as a hosted API on the StepFun platform; open weights are promised for October 15, 2026.

Alibaba's Qwen team released Qwen-Image-2.1 on September 20, a downloadable model that combines image generation, multi-image editing, and native transparent output in one package. It has a 7-billion-parameter visual generation component built from 32 single-stream diffusion transformer layers, generates at native 2K resolution, accepts up to 10 reference images, and can return RGBA files with alpha channels — deciding from the prompt whether an output needs transparency.

The release folds Qwen-Image-Layered's transparency handling into the main model. Consequential license catch: the weights are research-only by default; commercial use requires a separate license from Qwen. (UPDATE Sept 21: the weights had initially gone out under Apache 2.0 before Alibaba quietly switched them to the research-only terms.)

Chatter pointed to a Tuesday (Sept 22) release of GPT-6 Sol, a faster, cheaper everyday tier below GPT-6 Astra. The fuel: Sam Altman said his big launch slipped from last week to this week, and Tibo Sottiaux teased DevDay-level launches.

Fresh signals this morning suggest launch prep is actively underway. Tibo had promised a reset around 3 AM Pacific that hadn't landed by morning — though OpenAI routinely runs late on these things — and OpenAI employees are telling watchers they're "cooking," heads-down on something imminent.

The hardest evidence yet is an unverified reader tip: a screenshot shared with The AI Wire appears to show three new model IDs — gpt-6-sol, gpt-6-luna, and gpt-6-astra-minor — in the Azure OpenAI model registry. No public source for the listing has surfaced yet, so treat this as unconfirmed. Registry listings are a classic pre-launch tell — models typically surface there days before announcement. One correction from the registry: there is no "Aeon" entry; the third name is astra-minor, which supersedes earlier chatter about an Aeon bot.

Leak-based positioning has Sol at roughly 5x cheaper than Astra with near-flagship coding quality, aimed at being the daily driver below Astra. As of Tuesday morning nothing has actually shipped — OpenAI's catalog still shows only gpt-6-astra and gpt-6-astra-pro — but the registry sighting makes this the strongest the rumor has ever looked. It stays unresolved until OpenAI confirms or DevDay on Sept 29 passes without a launch.

UPDATE 2026-09-22 (Tuesday, 10:19 AM ET): no launch yet this morning — OpenAI's public catalog still lists only gpt-6-astra and gpt-6-astra-pro as GPT-6 identifiers, and the Thursday (Sept 17) launch date from last week's chatter passed with nothing shipping, per the cellcog.ai tracker. The only 'Sol' you can use today is still GPT-5.6 Sol (the July 9 release), so any GPT-6 Sol would be a second Sol in the catalog. We also checked Microsoft's public Azure Foundry region-availability doc — no gpt-6-sol or gpt-6-luna appears there yet, so the login-gated registry sighting from this morning remains the sole hard evidence, unconfirmed publicly. Tuesday is still young, and registry listings remain a classic pre-launch tell. If the day passes without a launch, DevDay on September 29 is the next date in play.

UPDATE 2026-09-22 (noon ET): the rumor mill is running hot ahead of a possible afternoon launch. Testing Catalog reports OpenAI teasing a 'major release on Tuesday' with product lead Tibo Sottiaux confirming the date on X — over the weekend he posted 'OK fine. But it's also still coming in Tuesday.' Builder chatter is converging on GPT-6 Sol and Luna (plus Terra) as the likely candidates.

More support from the user side: posters on OpenAI's own community forum (Sept 21–22) describe a 'SOL' tier appearing in their accounts, with complaints that Sol is a cheap thinker but an expensive 'doer' — anecdotal, but it reads like real users seeing a real model id. @angelbrodin captured the mood on X: 'I'm sol excited.'

On the public-evidence front, we re-checked Microsoft's public Azure Foundry region-availability doc today: still no gpt-6-sol or gpt-6-luna listed (only gpt-6-astra), so the login-gated registry sighting from this morning remains the only hard-but-unverified evidence. The pin stays. If OpenAI ships this afternoon, this entry flips to released; if Tuesday passes quietly, the pin comes down and DevDay (Sept 29) becomes the next date in play.

RESOLVED — CONFIRMED: OpenAI officially launched GPT-6 Sol (and Luna) on Tuesday, September 22, exactly the timing this rumor predicted. Reuters and 9to5Mac confirmed pricing and the Codex/ChatGPT rollout; see the release entry 2026-09-22-openai-gpt-6-sol-luna-released for full details.

The Windows and macOS Copilot apps are getting an embedded web-browser panel that sits side-by-side with the AI conversation. Browse and chat in the same window — Copilot can see what you’re looking at.

It’s opt-in and admin-gated: organizations enable it through the BrowsingEnabled policy, so IT shops worried about an AI reading browser tabs can keep it off. Broad rollout lands at the end of November.

The direction is clear — every major assistant wants to live next to the browser. Whether users keep the panel open is the open question.

AX (agentexecutor.io) is a Kubernetes-native control plane for AI agent workloads: declare tasks as YAML manifests, run them as sandboxed actors with suspend/resume, pre-wired workspaces, and network fencing — with a claimed ceiling of billions of concurrent tasks per cluster. It's Apache 2.0 at github.com/google/ax, built by the team behind Google's internal agent runtime, but explicitly pre-stable with breaking changes expected before any stable release.

A Reuters-sourced report citing three sources says Anthropic is considering rolling out a new AI model to counter OpenAI's GPT-6 Astra, which has reportedly gained traction among business users; timing is under discussion ahead of an expected IPO, and Anthropic declined to comment. No model name, capabilities, or launch schedule is confirmed.

The deliberation sits in tension with CEO Dario Amodei's September 12 essay calling for an industry-wide slowdown in capability releases — a stance that drew public support from Sam Altman and Elon Musk. Some investor sources say they are re-evaluating Anthropic's enterprise position as rising interest rates sharpen focus on profitable cash flows.

UPDATE 2026-09-21: the IPO framing is amplifying — Inspirepreneur Magazine (Sept 21) reports Anthropic is weighing the model launch as part of IPO preparation, citing the Daily Star's reporting that talks are ongoing while Anthropic weighs safety evaluation and the investment needed. New concrete datapoint behind the chatter: Ramp-tracked enterprise AI spend shows roughly 13% going to GPT-6 Astra versus ~8% to Claude Fable. Still no model name, capabilities, or launch date from Anthropic.

UPDATE 2026-09-22 — CONFIRMED: Anthropic officially launched Claude Opus 5.5 on Tuesday, its first new model since the Reuters report that it was weighing a launch to counter GPT-6 Astra. Opus 5.5 arrives with Fable-level performance claims and 20%-lower pricing; see the release entry 2026-09-22-anthropic-opus-5-5-released for full details.

Update (2026-10-08): Confirmed: Anthropic completed the Claude 5.5 family — Sonnet 5.5 and Opus 5.5 arrived first, and Claude Haiku 5.5 launched October 7 with up to 90% price cuts and a new adjustable effort setting.

President Trump announced plans to appoint a new AI czar and create an 'AI Force' in a series of Truth Social posts, vowing to keep the US ahead of China in AI without slowing the industry. He dismissed safety concerns as a 'hoax,' writing that the administration 'will not in any way hinder or stifle the Growth of this incredible Industry,' and said it would address AI harms through existing criminal and civil law rather than new sector-specific regulation.

The announcement lands as a direct counter-framing to the month's safety cascade. It came within days of all four frontier lab chiefs publicly agreeing to pace development and accept embedded evaluators, plus California's first-in-the-nation AI safeguards bill. Trump also framed industry calls to slow down as a 'Market Fixing' ploy by figures like Anthropic's Dario Amodei and OpenAI's Sam Altman.

Concrete details are thin so far. As of mid-September there is no executive order establishing the AI Force, no legislation, no named czar — the role was previously held by David Sacks, who stepped down in the spring — and no stated budget, host department or remit. Axios reports the AI Force is modeled on the Space Force; with that much still undefined, this one stays 'announced' until the machinery shows up.

Alibaba's Qwen team released Qwen3.8-LiveTranslate on September 19, a real-time simultaneous interpretation model covering 60 languages, with average lag down to 2.3 seconds (an ~18% cut from 2.8s) on the LAAL metric. It listens to live speech, optionally processes video frames, and returns translated text and speech while the speaker is still talking; it's available now as a hosted API on Alibaba Cloud Model Studio and QwenCloud under model ID qwen3.8-livetranslate-flash-realtime over WebSocket, with a Flash variant for offline audio/video translation.

Single-source so far (Complete AI Training); Alibaba's launch announcement is reportedly still marked draft while production docs list the model as available.

Musk told a podcast that Grok 4.7 would ship 'next week,' but that window passed without a release. Musk instead announced Grok 4.8 on September 13 (2.5T parameters, new C++ stack) and downgraded Grok 4.7 to 'roughly comparable to Anthropic's Opus 5.0' — xAI has still published no model card, price, or release date for 4.7, and attention has moved to 4.8.

UPDATE 2026-09-22: confirmed — Grok 4.7 launched on September 21 for coding and knowledge work, after weeks of delays (see the 2026-09-21 Grok 4.7 released entry). This rumor resolved exactly on schedule.

AI creator Matthew Berman's latest upload is a sit-down with OpenAI's Tibo Sottiaux — the face behind the ChatGPT usage-reset announcements — on how to read the next wave of AI developments. Second Berman video on Tibo this month after his earlier Pragmatic Engineer coverage; Tibo's posts are also central to the GPT-6 Sol rumor cycle.

SpaceXAI (xAI) released Grok Voice Transcribe 2.0 on September 18, a speech-to-text model it calls the most accurate available — claiming double the accuracy of version 1.0 at unchanged pricing: $0.10 per audio hour for batch transcription and $0.20 for real-time streaming, with speaker diarization, word-level timestamps, and key-term biasing included. It is built on the audio foundation behind Grok Voice, trained on live, noisy, multilingual audio and refined with post-training.

The company says word error rate fell from 20.6% to 6.8% on short multilingual voice commands, from 10.6% to 7.1% on telephony, and from 8.7% to 3.3% on conversational audio — and that it ranks first among 32 streaming models on the public Artificial Analysis leaderboard. Those are vendor-reported numbers; independent confirmation matters more than usual here, since xAI's Terminal-Bench claims on Grok 4.7 recently showed a 12-point gap to independent runs.

Distribution is the interesting part: Atlassian's Loom now uses it to transcribe every video, per RuntimeWire. The model is live through the Speech-to-Text API as grok-voice-transcribe-2.0, with 1.0 headed for deprecation in the coming weeks — developers who need the old model should pin grok-voice-transcribe-1.0 explicitly.

Qwen's first omni-modal model accepts text, images, audio, and video with 1M context and is built for agentic tool-calling workflows. API is $0.15 in / $0.47 out per million tokens, with audio input pricing cut roughly 98%. Live on QwenCloud, Alibaba Cloud Model Studio, and Qwen Studio; no open weights announced.

Researchers at AIR disclosed Plugin4Shell, a zero-click remote code execution vulnerability shared by four major AI coding agents — Claude Code, OpenAI's Codex, GitHub Copilot and Google's Gemini CLI. All four pin plugins to a reviewed commit hash but check out the pinned commit without verifying the checkout landed there, letting an attacker swap in malicious code while the pin still looks intact. AIR calls it the first supply chain vulnerability of the AI agent ecosystem.

The flaw is zero-click because the same git checkout re-runs on background auto-update: when a marketplace bumps the pinned SHA, the swap reaches already-installed plugins with no user action. AIR says it previously demonstrated plugins spreading to more than 26,000 agents and found 925 hijacked skills in active use reaching 134,000 agents, proving the takeover chain end to end.

Vendor response was uneven. Anthropic patched Claude Code in version 2.1.179 and OpenAI patched Codex in 0.146.0. Microsoft has not shipped a fix for Copilot, leaving its users with no patch. Google deprecated Gemini CLI entirely rather than patching it, meaning existing installs stay exposed indefinitely; its advice is to move to its newer agent, Antigravity.

OpenAI announced Astra for Law on September 18, a legal-focused version of GPT-6 Astra for lawyers and legal-tech companies. It pairs the model with a dedicated legal search index spanning more than 230 million URLs of US case law, statutes, regulations, court rules and administrative decisions, plus custom instructions for legal analysis and 26 ecosystem plugins connecting specialist legal tools.

OpenAI says it hit 54% correctness on 200 Vals AI Legal Research Bench questions versus 38.7% for plain GPT-6 Astra with web search. Initially available to selected US law firms through Trusted Access — announced, not broadly released, consistent with the site's restricted-tier convention.