15 labs, 22 models. What each lab said at launch, next to what someone other than the lab found.
San Francisco, California
The company that put a chatbot in a billion pockets, now betting the balance sheet that it can out-spend its own compute bill before revenue catches up.
GPT-6 Sol coding · 22 Sept 2026
They saidComparing GPT-6 Sol to Anthropic's Claude Opus 5, GPT-6 scores 6.3% better at only 9% of the cost per task.
Independent testsNothing independent yet. Treat the launch numbers as marketing until someone else runs them.
ZDNET ↗GPT-6 Luna frontier LLM · 22 Sept 2026
They saidOpenAI reports that GPT-6 Luna beat its earlier GPT-5.6 release by 5.4%. The big news is that it also costs 58% less per task.
Independent testsNothing independent yet. Treat the launch numbers as marketing until someone else runs them.
ZDNET ↗MoneyOpenAI reportedly projects $280 billion in cash burn through 2030, even as annual revenue is forecast to rise from $36 billion in 2026 to $350 billion in 2030. TechRepublic ↗
Latest · 23 Sept 2026OpenAI said in an August blog post that it paused some reinforcement-learning training runs, strengthened monitoring and delayed a planned frontier-model training run pending further evaluation. Mint ↗
Check this nextWhether the $280 billion cash-burn forecast changes the IPO timeline, and whether GPT-6 Sol's self-reported cost and safety numbers hold up once outside evaluators get access.
San Francisco, California
The safety-first lab whose CEO asks the industry to slow down in one blog post and ships a new frontier model days later.
Claude Opus 5.5 frontier LLM · 23 Sept 2026
They saidAnthropic claimed Opus 5.5 produces easier-to-understand writing and "attempts to bypass constraints about 85 percent less often" than the previous model.
Independent testsNothing independent yet. Treat the launch numbers as marketing until someone else runs them.
Tempo.co ↗Claude Opus 5.5 frontier LLM · 23 Sept 2026
They saidAnthropic priced the model at $4 per million input tokens and $20 per million output tokens.
Independent testsThe company said the model delivers performance comparable to Claude Fable 5.1 on most tasks while costing 40% less to operate than Opus 5.1.
Mint ↗MoneyAnthropic is moving ahead with plans for a November IPO that could value the Claude developer at about $2 trillion and raise as much as $100 billion. CoinGape ↗
Latest · 23 Sept 2026Anthropic expanded its Claude Marketplace to more than 2,000 connectors, plugins, agents and products from partners including Cursor, CrowdStrike, Accenture and Deloitte. Crypto Briefing ↗
Check this nextWhether the $2 trillion IPO valuation survives Dario Amodei's own public call for a slower pace of frontier development, and whether the 85%-lower bypass-attempt rate holds under outside red-teaming.
London, United Kingdom
Google's research lab merged with product urgency, shipping a Gemini variant almost every month while its own leadership chart gets rewritten underneath it.
Gemini 3.8 Live Extended Thinking reasoning · 15 Sept 2026
They saidGemini 3.8 Live Extended Thinking, the higher-effort model, beats rivals on quality for $3.50 an hour.
Independent testsThat compares to $4.80 an hour for Grok Voice Think Fast 2.0, per an independent Artificial Analysis cost benchmark cited in the same report.
OfficeChai ↗MoneyCEO Sundar Pichai said Google Cloud revenue grew 82% year over year in the second quarter of 2026, and its backlog hit $514 billion. The Motley Fool ↗
Latest · 5 Aug 2026Alphabet made Demis Hassabis Chair of Google DeepMind and Chief Scientist of Alphabet, handing day-to-day control of DeepMind to Koray Kavukcuoglu, as Gemini passed 950 million monthly users. Free Press Journal ↗
Check this nextWhether the DeepMind leadership split (Hassabis on AGI/science, Kavukcuoglu on Gemini shipping) speeds up releases or just adds a layer of management between them, and whether Gemini 3.5 Pro ever actually ships.
Menlo Park, California
The open-weights evangelist that also runs a Superintelligence Labs talent war and, on its own account, built its newest agent by studying an open-source project.
Muse frontier LLM · 8 Sept 2026
They saidMeta maintains, however, that no code was directly copied and that Muse was built from the ground up.
Independent testsMuse was "heavily inspired" by OpenClaw, an open-source, self-hosted agent framework, and the similarities extend to workspace file structures and agent architectures.
Crypto Briefing ↗MoneyMeta's FY2026 capex guidance came in between $130 billion and $145 billion to fund AI infrastructure, after capex hit $31.1 billion in the most recent quarter. 24/7 Wall St. ↗
Latest · 23 Sept 2026Muse, Meta's personal AI agent, racked up over 900,000 downloads in its first week on the market. Crypto Briefing ↗
Check this nextWhether Meta discloses more about how closely Muse's architecture tracks OpenClaw's as scrutiny builds, and whether the $130-145 billion capex guide holds if Muse's download surge fades.
Bastrop, Texas
Elon Musk's frontier lab, which measures progress in parameter counts and promised release dates, and misses the second one almost every time.
Grok 4.7 coding · 21 Sept 2026
They saidGrok 4.7 packs 2.1 trillion parameters, up 40% from the 1.5 trillion in Grok 4.6, itself a refinement of Grok 4.5.
Independent testsNothing independent yet. Treat the launch numbers as marketing until someone else runs them.
Decrypt ↗MoneyxAI raised $20 billion in its latest funding round, surpassing its initial $15 billion target, made up of $7.5 billion in equity and $12.5 billion in Nvidia-GPU-backed debt. Proactive Investors ↗
Latest · 20 Sept 2026xAI released Grok 4.7 after at least five delays since late July. Decrypt ↗
Check this nextWhether Grok 4.7's self-reported 2.1-trillion-parameter jump translates into independent benchmark wins, and whether the next release date Musk gives holds any better than the last five did.
Redmond, Washington
Owns a chunk of OpenAI and still builds its own MAI models, in case renting frontier AI from a partner stops being a plan.
MAI-Cyber-1-Flash cybersecurity · 27 July 2026
They saidMicrosoft says MDASH running MAI-Cyber-1-Flash alongside GPT-5.4 scores 95.95% on CyberGym, beating Anthropic's Mythos and other cyber models, up from 88.45% before the new model joined the harness.
Independent testsThe headline 95.95% is not the small model on its own; it is the MDASH harness running MAI-Cyber-1-Flash alongside GPT-5.4, up from 88.45% in May 2026.
MarkTechPost ↗MAI-Transcribe-2 audio · 3 Sept 2026
They saidThe limited-time $0.10-per-hour rate is about 72% lower than the $0.36-per-hour price listed for MAI-Transcribe-1 and 1.5.
Independent testsArtificial Analysis currently ranks the model second overall for accuracy, behind Alibaba's streaming Fun-Realtime-ASR preview, and Microsoft Learn lists it as a preview with no SLA, not recommended for production.
eWeek ↗MoneyMicrosoft's cloud platform Azure surpassed $100 billion in annual revenue for the first time, in fiscal Q4 2026, the quarter that ended June 30. The Motley Fool ↗
Latest · 14 Sept 2026Microsoft published a draft Humanist AI Code of Conduct for its AI model development, open for public feedback for six weeks before a revised version ships. SSBCrack News ↗
Check this nextWhether MAI-Transcribe-2's 10-cent rate survives past its end-of-2026 expiry, and whether Copilot's default model ever becomes an in-house MAI instead of OpenAI's GPT line.
Seattle, Washington
Runs the world's biggest cloud and is now betting its agent stack, and Alexa's voice, on models it builds in-house rather than rents.
Nova 2 frontier LLM · 9 Dec 2025
They saidAmazon positions Nova 2 as a frontier-grade model tightly integrated with Amazon Bedrock and its new AgentCore framework, unveiled at AWS re:Invent 2025.
Independent testsOne cloud-architecture columnist called adopting Nova 2 a strategic risk rather than a technical one: it anchors agentic workflows in APIs, runtimes and orchestration semantics that exist only inside AWS.
InfoWorld ↗Alexa+ (India launch) assistant · 16 Sept 2026
They saidAmazon says users in India interacted with Alexa more than nine billion times over the past 12 months, ahead of bringing the generative AI assistant Alexa+ to the country.
Independent testsAmazon has not specified how data from Alexa+ interactions is used to train its AI models, nor said when the free Early Access period will end.
MediaNama ↗MoneyAWS generated $42.2 billion in revenue in the second quarter of 2026, up 37 percent year over year, its fastest growth in years. ad-hoc-news.de ↗
Latest · 28 July 2026Amazon has begun deprecating most of its in-house flagship Nova models, including the Premier and Omni models, the Reel video model and the Canvas image model, shifting resources to a new frontier model effort led by researcher Pieter Abbeel. Business Insider ↗
Check this nextWhether the new frontier model that Pieter Abbeel's team is building actually ships at this year's re:Invent, or whether Amazon quietly slips the deadline again.
Santa Clara, California
Sells the chips every AI lab is built on, and is now shipping its own open models too, in case demand for anything other than Nvidia GPUs ever mattered.
Nemotron 3 Diarization open-weights · 23 Sept 2026
They saidNVIDIA says Nemotron 3 Diarization scored 14.72% DER against 19.3% for the next-ranked system on Voice Arena's initial Diarization-Bench, a roughly 24% relative reduction, ranking first among rival systems.
Independent testsNVIDIA notes the results may change once Voice Arena completes its official Version 1 evaluation, and discloses one regression of its own on a two-speaker test even as headline diarization scores improved elsewhere.
MarkTechPost ↗MoneyNVIDIA reported revenue of $96.2 billion for the quarter ended July 26, 2026, up 18 percent from the previous quarter and up 106 percent from a year ago. Yahoo Finance (GlobeNewswire) ↗
Latest · 26 Aug 2026NVIDIA announced strategic partnerships with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to mobilize over $500 billion of third-party capital for AI infrastructure, subject to definitive agreements. Yahoo Finance (GlobeNewswire) ↗
Check this nextWhether the $500 billion in financing partnerships turns into signed deals, or stays a subject-to-definitive-agreements press line.
Cupertino, California
Sells the hardware everyone already trusts and is racing to bolt a credible AI assistant onto it, three years after first promising one.
Siri AI (Apple Foundation Models) assistant · 9 June 2026
They saidApple says the new Siri AI runs on its own family of five Apple Foundation Models rather than on Google's Gemini, even though Gemini's outputs helped train it, engineering VP Craig Federighi said at WWDC 2026.
Independent testsTo bring its most demanding cloud model to life, Apple harnessed the power of Nvidia GPUs housed in Google's own cloud servers to scale its Private Cloud Compute architecture.
PCMag ↗MoneyApple reported total revenue of $109.4 billion for its June 2026 quarter, up 16 percent, a new record for that quarter. Yahoo Finance (Variety) ↗
Latest · 30 July 2026Tim Cook will step down as Apple's CEO after 15 years, with hardware chief John Ternus succeeding him and Cook becoming executive chairman. Yahoo Finance (Variety) ↗
Check this nextWhether the new Siri AI, trained partly on Gemini's outputs and demoed at WWDC 2026, actually ships broadly on iOS 27 without another year-long delay like the last one.
Paris, France
Europe's flagship open-weight lab, betting that giving models away can still build a business, while its own government worries Europe bet too much on one company.
Voxtral TTS audio · 26 Mar 2026
They saidMistral says Voxtral TTS delivers 70ms model latency for a 10-second voice sample and 500 characters of text, with a real-time factor of up to 9.7x, and supports nine languages.
Independent testsMistral is pitching an open-weight, smaller model against established closed players like ElevenLabs and OpenAI's Speech API, but ships it under a noncommercial license that limits real production use.
Yahoo News Australia ↗Robostral Navigate robotics · 8 July 2026
They saidMistral says Robostral Navigate enables robot navigation using a single camera, without lidar or multiple-camera setups, and is designed to work across robots from different suppliers.
Independent testsNothing independent yet. Treat the launch numbers as marketing until someone else runs them.
Yahoo News Australia (Reuters) ↗MoneyMistral raised 3 billion euros in a Series D round led by Samsung Electronics, at a post-money valuation of more than 21 billion euros, the largest equity round ever completed by a European tech company. HPCwire ↗
Latest · 3 Sept 2026France's finance minister Roland Lescure warned that if European AI is just about Mistral, then Europe is doomed, calling instead for a whole ecosystem, even while praising Mistral as a company Europe can be proud of. Malay Mail (AFP) ↗
Check this nextWhether the new 21 billion euro valuation survives contact with the French finance minister's own warning that Europe can't build an AI ecosystem around one company.
Hangzhou
Hangzhou quant fund's AI spinoff, still proving frontier performance does not require frontier compute or Nvidia.
DeepSeek V4.1 Flash open-weights · 10 Sept 2026
They said552-billion-parameter model built to outperform key rivals while costing dramatically less to run.
Independent testsNothing independent yet. Treat the launch numbers as marketing until someone else runs them.
Crypto Briefing ↗DeepSeek V4 Pro (GA) reasoning · 13 Aug 2026
They saidGeneral-availability V4 Pro scores 62.7 on the DeepSWE coding benchmark, up from 12.8 in the April preview.
Independent testsIndependent evaluator Artificial Analysis ranks it below Moonshot's Kimi K3 and Anthropic's Claude Opus 5 on its Intelligence Index despite the price gap.
eWeek ↗MoneyDeepSeek hired CITIC Securities to prepare for a potential IPO on Shanghai's STAR Market as it seeks capital to expand compute. Tekedia ↗
Latest · 21 Sept 2026DeepSeek plans to deploy at least 160,000 Huawei Ascend 950DT chips at a data center in Inner Mongolia, about 1 gigawatt in scale, for inference. Startup Fortune ↗
Check this nextWhether the Huawei Ascend cluster in Inner Mongolia actually holds up running inference at scale, or DeepSeek quietly routes traffic back to Nvidia.
Hangzhou
Ecommerce giant betting its next decade on Qwen and 20 gigawatts of cloud compute instead of shopping carts.
Qwen3.8-Max frontier LLM · 1 Aug 2026
They saidAlibaba calls it the strongest model in the Qwen series to date.
Independent testsOn the Text Arena leaderboard it ranks fifth, behind several models built by Anthropic.
Proactive Investors ↗Qwen 4 frontier LLM · 23 Sept 2026
They saidQwen 4 is in training now, with Qwen 4.5 and Qwen 5 roadmapped to scale up to 5 to 10 trillion parameters.
Independent testsNothing independent yet. Treat the launch numbers as marketing until someone else runs them.
Vietnam Investment Review ↗MoneyAlibaba raised $10.2 billion in a discounted Hong Kong share sale, the largest follow-on offering on record, to fund AI infrastructure. Memeburn ↗
Latest · 19 Sept 2026Alibaba's Qwen team launched Qwen3.8-Omni-Flash on September 18, cutting audio API pricing more than 98% and beating its predecessor by 26% across 30 benchmarks, but API-only with no open weights. Startup Fortune ↗
Check this nextWhether Qwen 4 actually lands near 10 trillion parameters, or whether China's ongoing AI data-security probe into Alibaba slows the roadmap it just announced.
Beijing
Beijing lab behind the Kimi assistant, betting a single giant open-weight model can out-cheap the US frontier labs into an IPO.
Kimi K3 open-weights · 17 July 2026
They saidMoonshot AI's Kimi K3, an open-weight model its developer describes as the first open model to reach 2.8 trillion parameters, became available on Amazon Bedrock on September 18, 2026, adding a new option for coding and knowledge work.
Independent testsNothing independent yet. Treat the launch numbers as marketing until someone else runs them.
Unite.AI ↗MoneyMoonshot AI is raising new capital at a $50 billion valuation, in line with rival Z.ai, and could pursue a Hong Kong IPO before the end of 2026. Briefs ↗
Latest · 23 Sept 2026Anthropic alleges Moonshot routed real Kimi user requests, including one about surveillance cameras near Chinese military sites, to Claude without users' knowledge, and China's cyberspace regulator has opened a probe. Ynetnews ↗
Check this nextWhether China's cyberspace-regulator probe into the Claude-routing allegations derails Moonshot's $50 billion raise and Hong Kong IPO plans.
Beijing
Tsinghua University spinoff and the first pure-play LLM company to go public, betting domestic chips can match Nvidia on cost per token.
GLM-5.3-Flash open-weights · 27 Aug 2026
They saidZhipu says its newest Flash model, the first native multimodal release in the GLM series, runs its inference service entirely on domestically made chips instead of Nvidia GPUs.
Independent testsIn benchmark testing, the model scored 57 points on the Artificial Analysis Intelligence Index, on par with Anthropic's Claude Opus 4.8.
Global Times ↗MoneyZhipu has raised nearly $10 billion in public markets since its January 2026 Hong Kong IPO, including a $4 billion share placement in July and a $5 billion round in September. Crypto Briefing ↗
Latest · 17 Sept 2026Zhipu says an AI agent optimized its own GLM-5.3-Flash inference service running on more than 100,000 domestic chips, taking over all online traffic within two weeks and roughly tripling end-to-end throughput. 36Kr ↗
Check this nextWhether Zhipu's self-optimizing 'Infra Agent' is genuine recursive self-improvement or a one-off engineering stunt dressed up as an RSI milestone.
San Francisco
Mira Murati's ex-OpenAI crew, betting that giving customers a model they can fine-tune themselves beats selling one-size-fits-all chatbots.
Inkling open-weights · 15 July 2026
They saidInkling packs 975 billion parameters into a sparse Mixture-of-Experts design, with roughly 41 billion active on any given inference pass.
Independent testsIt leads every US open-weights model but trails China's best on the Artificial Analysis Intelligence Index.
Memeburn ↗MoneyThinking Machines Lab is in talks to raise $5 billion to $6 billion at a pre-money valuation of at least $40 billion, with Nvidia supplying roughly half the capital. The Next Web ↗
Latest · 23 Sept 2026Thinking Machines signed a $65 million annual deal with Crusoe Cloud on September 23, 2026, to run production inference for Inkling and its fine-tuned variants on dedicated Nvidia HGX B200 clusters. Unite.AI ↗
Check this nextWhether the $5-6 billion raise at a $40 billion pre-money valuation actually closes, given a string of founder departures and self-reported, unaudited revenue in only the hundreds of millions.