Wednesday, 23 September 2026

OpenAI released GPT-6 Sol and GPT-6 Luna, the two smaller models in its GPT-6 family, and cut their API prices by 50% against the GPT-5.6 promotional pricing [1][2]. GPT-6 Sol costs $2 per million input tokens and $10 per million output tokens, down from $4 and $20. GPT-6 Luna costs $0.10 and $0.50, down from $0.20 and $1.20 [1]. The company says improved prompt caching now gives higher cache hit rates by default, with a 90% discount on cached input reads [1].

OpenAI says Sol makes about half as many mistakes as its predecessor on an internal factuality test built from conversations where users flagged errors [1]. On the DeepSWE coding benchmark it reports Sol at 68.8%, against 69.9% for Anthropic's Claude Fable 5 at xhigh effort, at about 80% lower cost per task [1]. These are vendor numbers, run by the vendor.

Both models are in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu users; Free and Go users get Luna in the desktop app, and neither model is in Chat yet [1]. In the API they are `gpt-6-sol` and `gpt-6-luna` [1]. TechCrunch notes Anthropic released Opus 5.5 about 90 minutes earlier [2]. Simon Willison points out that GPT-5.6 has a 25% price increase scheduled for November, so the comparison is against promotional rates, and that GPT-6 Sol now matches GPT-5.6 Terra on input price [3].

Google released two text-to-speech models on 23 September, Gemini 3.8 Flash TTS and Gemini 3.8 Flash-Lite TTS, rolling out in Google AI Studio and the Gemini API [1][2]. Flash TTS is pitched at creative work such as games, audiobooks and podcasts; Flash-Lite TTS is the high-volume, lower-cost option aimed at dubbing and voice agents [1][2].

With Flash TTS, a user can describe a voice in plain language, including role and accent, and generate it across more than 100 languages and dialects [1]. There is also a library of more than 2,000 ready-made voices [1]. The feature most operators will notice is replication: Flash TTS can recreate a voice from a 30-second audio sample [1]. Google says the user must supply a verbal consent recording from the voice owner, and the system checks it matches the reference speaker before the voice is created [1]. Every clip carries a SynthID watermark, and replicated voices also carry C2PA content credentials [1][2].

Both models take line-by-line stage directions, handle two-speaker scenes from one script, and support cues such as laughs and sighs [1]. Google says Flash TTS placed first on Hume AI's Voice Design Benchmark with a score of 71.4 [1]. It names Figma, HeyGen and Wondercraft among companies integrating the models [2]. Pricing is not stated in the announcement, and TNW reports that enterprise access through Gemini Enterprise is coming soon [2].

OpenAI is bringing voice-driven agent tasks to the ChatGPT mobile app, TechCrunch reports [1]. Plus and Pro users can use the Work tab on their phone to create a document, draft an email or summarise Slack messages, and can build sites, create presentations and use the cloud browser [1]. Free and Go users get plugins and connected apps [1].

9to5Mac lists three changes to ChatGPT Voice in the same update [2]. Voice can now run on OpenAI's three GPT-6 models, Astra, Sol and Luna. Plugins, including email, calendar and Slack, now work inside Voice. And Voice now works with ChatGPT Work on the web and mobile [2]; TechCrunch reports it previously did so only in the desktop app [1]. According to TechCrunch, voice conversations will show richer text output, users can switch between voice and text, and a conversation started on a phone can be resumed on desktop [1].

The practical shift is that the same agent that reads your inbox or drafts a document can now be started by speaking to a phone. That makes connected-app permissions a mobile question as well as a desktop one: whatever ChatGPT Work can touch at your desk, it can now touch from a voice request in a car or a hallway [1][2]. OpenAI first added voice control to the desktop app in July, according to 9to5Mac [2].

The UN Security Council holds a high-level briefing on AI and international security on 23 September, convened by France during the General Assembly's high-level week and chaired by French Foreign Minister Jean-Noël Barrot [1][2]. The briefers are Yoshua Bengio, co-chair of the UN's Independent International Scientific Panel on AI, OpenAI CEO Sam Altman, Anthropic CEO Dario Amodei and Hugging Face's Clément Delangue [1][2].

Security Council Report says it is the Council's first meeting focused specifically on the safety risks of increasingly capable AI systems, including loss of human control [1]. France's concept note, it reports, frames the session around misalignment and points to recursive self-improvement, where AI systems develop more capable successors, as a further risk [1]. According to the Reuters report carried by BNN Bloomberg, an OpenAI executive said Altman would urge world leaders to adopt benchmarks for measuring AI capabilities and assessing the safeguards companies put in place [2]. Security Council Report says OpenAI has described such standards as a common technical foundation rather than mandatory pre-release approval [1].

The politics are split. Donald Trump told the General Assembly on 22 September that he "rejects any attempt to construct a globalist scheme to control" AI, according to Reuters [2]. Reuters also reports that the US and China have discussed a notification system that would cover AI incidents at a national-security level [2]. For businesses, the useful signal is which evaluation and oversight standards the labs are asking governments to adopt [1][2].

Snorkel AI raised $350 million in a Series E at a $3.5 billion valuation, co-led by Insight Partners and S32 [1][2]. TechCrunch reports the valuation is nearly triple the $1.3 billion set when Snorkel raised $100 million in a Series D 17 months ago [1]. Existing investors including Addition, Greylock, Lightspeed and GV took part [1][2].

The business has changed shape. Snorkel started as software for automating data labelling and shifted last year to selling finished datasets, which it calls data-as-a-service, according to TechCrunch [1]. The company says it now builds expert tasks, simulated environments and grading rubrics for AI labs and enterprises, the material used to train and evaluate agentic models [2]. Snorkel says its annualised revenue run rate is $375 million, an eighteenfold increase in 12 months [1]. Because Snorkel sells environments and datasets rather than expert labour, the company says, payments to its human experts sit in cost of goods sold rather than in that revenue figure; TechCrunch notes that labour marketplaces such as Mercor report gross figures while paying roughly 60% to 70% of top-line income to their experts [1].

The company says it will use the money to expand what it calls its agentic data factory, invest in vertical and enterprise AI, and extend its research into new domains [2]. For teams building their own agents, the round is a price signal: good evaluation and training data is now a product that labs pay heavily for [1][2].

Ema, a Mountain View startup that runs teams of AI agents across HR, IT and finance work, raised a $77 million Series B led by Bengaluru-based Creaegis, with Accel, S32 (Section 32) and Prosus adding to their stakes [1][2]. The round takes total funding to $140 million and more than quadruples the valuation set in 2024; the company did not disclose the new figure [1][2].

Ema's pitch is that its agents wrap around a company's existing applications and carry out multi-step processes, then let the customer lean less on some of that software over time [1]. CEO Surojit Chatterjee told TechCrunch the product can draw on more than 150 models and that the company focuses on domain knowledge, integrations and orchestration rather than competing with frontier labs [1]. The company says its platform connects to more than 250 business applications and that customers include Wipro, Hitachi, ADP and PwC [2].

The numbers are company claims. Ema says it has more than 50 active enterprise deals and over 1 million active users, that revenue has grown 50-fold in two years, and that bookings have passed $150 million, a figure Chatterjee said counts the full value of multiyear contracts rather than annual revenue [1]. He put net dollar retention near 180% [1]. Ema does not price by seat or by token; it charges on completed tasks and outcomes, according to TechCrunch [1]. Much of the money goes to sales and marketing, and the company plans to expand into Asia-Pacific, South America and parts of the Middle East over the next year [1].

Google DeepMind published a technical update on 23 September describing how it will add persistent, cross-device memory to Private AI Compute, the cloud platform Google introduced to run Gemini models on personal data with on-device-style privacy [1][2]. Until now, the company says, Private AI Compute and similar systems across the industry were stateless: they wiped all context when a task ended [1].

Under the new design, the information an assistant needs is stored in dedicated encrypted storage in the cloud, while the keys to unlock it are held only on the user's own devices, which Google says makes the data inaccessible to anyone else, including Google [1]. When a model needs that context, an end-to-end encrypted channel connects the device to an isolated secure enclave that decrypts the data in memory, handles the request, saves any new context and re-encrypts it [1]. Google's original announcement says the platform runs on its own TPUs inside what it calls Titanium Intelligence Enclaves [2].

Google gives examples such as resuming on a laptop what a user viewed through smart glasses, or carrying a conversation between mobile and web [1]. It says it is publishing an updated whitepaper, a tamper-proof public record of its server software that devices can check before sending personal data, and the results of an independent audit by a cybersecurity firm it does not name in the post [1]. Which products will use the memory layer, and when, is not public [1].