This week, frontier models split into price tiers and the money behind them went public
OpenAI, Google and Microsoft shipped mid-priced models and leaner agent plumbing, while Anthropic closed a $93.5 billion IPO and NVIDIA reported a $115.8 billion quarter.
Covers 2026-11-15 – 2026-11-21
The week of November 15 had two stories running side by side. On the product side, the big labs kept splitting their lineups into price tiers and shipping plumbing that lets agents carry less context per call. On the money side, the companies building and supplying those models put numbers on the table: Anthropic completed the largest IPO by proceeds on record, according to TechCrunch, and NVIDIA reported a $115.8 billion quarter. Taken together, frontier AI is now sold by the task and financed in public markets.
Model choice is now a price-tier decision
Three vendors released mid-priced models within two days. OpenAI shipped GPT-6.2 Astra at $8 input and $40 output per million tokens, which The Decoder reports is 20% below its predecessor, and GPT-6.2 Sol at an unchanged $2/$10. Both have a 1,050,000-token context window, per OpenAI's model pages. The same changelog entry added a public beta of two-hour and eight-hour background completion windows in the Responses API, priced 30% and 40% below Standard rates. Work that can wait now has a cheaper rate.
Google released Gemini 4 Neon at $3/$15 with a 1 million-token window (announcement). That places it between Hydrogen ($1.50/$7.50) and Carbon's standard $4/$20; Carbon is currently on an introductory $2/$10. Google says Neon's cost per task is 32% lower than Carbon's at its default thinking level. That is Google's own figure.
At Ignite on November 17, Microsoft announced its own coding models, Code 2 and Code 2 Flash, generally available in GitHub Copilot and Foundry. It described them as frontier-level on cost-performance for enterprise coding but gave no independent evaluation. It also announced Microsoft-Decision-1.1, brought GPT-6.2 to Foundry, and introduced "MAI Licensed Weights" for trusted customers. We could not open independent coverage of the Ignite models, so those claims rest on Microsoft's post alone.
For developers, the vendors now present cost per completed task as the main metric, not price per token. Every one of those cost-per-task figures comes from the vendor. Before you switch a production workload, run your own evaluation on the cheaper tier.
Agents got leaner plumbing, with some new constraints
- Google Tool Retrieval (preview, Interactions API): tools marked for deferred loading send only names and descriptions, and the model fetches a full schema when it needs one. Google says this reduces input tokens and latency but published no figures (post).
- Hugging Face moved Inference Providers onto OpenRouter's router for language models. The router picks a provider per call on live price, speed and reliability and retries on failure. Hugging Face says code stays the same, but custom provider API keys no longer work for language models (post).
- OpenAI tied GPT-6.2's protected reasoning to the conversation that produced it. If you rewrite an earlier message, the API drops the reasoning that followed it, and older model families cannot read GPT-6.2's reasoning. OpenAI's distillation post reports that in its own red-teaming, the success rate for extracting reasoning from an earlier request fell from 54.0% to 0.2%.
For agent builders the practical change is the same across all three: keep conversation prefixes stable and load tools only when they are needed, or you lose cached reasoning and cache hits.
The money went public
Anthropic priced 980 million Class A shares at $83 on November 18, above its $70 to $78 range, and began trading on Nasdaq as ANT on November 19. Its November 20 closing notice says the full underwriter option brought the total to 1,127,000,000 shares and about $93.5 billion in gross proceeds, roughly $91.6 billion of it to Anthropic. A concurrent private placement of about $26.0 billion went to AMD, Broadcom, Google, Micron, Microsoft, NVIDIA and TSMC Partners. TechCrunch put the fully diluted valuation near $1.5 trillion.
NVIDIA reported third-quarter revenue of $115.8 billion, up 103% year over year, with $107.6 billion from data centers. It guided to $130.0 billion, plus or minus 3%, for a fourth quarter that includes an extra week, and it assumes no data center compute revenue from China. This was its first report under interim CEO Colette Kress. The same day, its board created a standing national security and export control oversight committee. CNBC reported the shares had fallen for four straight sessions going into the report.
The throughline
Model prices are being cut and tiered at the same moment that the capital behind the models has grown enormous and moved into public markets. Those two facts do not conflict: the chip and cloud suppliers in Anthropic's private placement are paid by the same compute bill that cheaper per-task pricing is designed to grow. For users, the useful takeaway is narrower: more mid-priced options, cheaper deferred processing and leaner tool loading, each still backed mainly by the vendor's own numbers.
Why this matters
- Model selection is increasingly a choice between price tiers, and the vendors' cost-per-task claims still need your own testing.
- Agent builders face new rules: stable conversation prefixes and deferred tool loading now affect both cost and whether reasoning is preserved.
- The capital behind frontier AI is now public, with chip and cloud suppliers holding stakes in a leading model maker.
Key takeaways
- GPT-6.2 Astra costs $8/$40 per million tokens, and Sol stays at $2/$10; background completion windows are 30% to 40% cheaper.
- Gemini 4 Neon costs $3/$15 with a 1 million-token context window.
- Microsoft's Code 2 and Code 2 Flash are generally available in GitHub Copilot and Foundry, with no independent evaluation yet.
- Anthropic's IPO closed at 1,127,000,000 shares and about $93.5 billion gross, plus a $26.0 billion private placement.
- NVIDIA reported $115.8 billion in revenue and guided to $130.0 billion, assuming no China data center compute revenue.
Sources
- OpenAIPrimaryGPT-6.2 Astra model pagedevelopers.openai.com
- OpenAIPrimaryGPT-6.2 Sol model pagedevelopers.openai.com
- OpenAIPrimaryOpenAI API changelogdevelopers.openai.com
- OpenAIPrimaryCombatting adversarial distillationopenai.com
- GooglePrimaryIntroducing Gemini 4 Neonblog.google
- GooglePrimaryIntroducing Tool Retrieval in the Gemini APIblog.google
- Hugging FacePrimaryHugging Face and OpenRouter join forces to make open-weight models easy to runhuggingface.co
- MicrosoftPrimaryIgnite 2026: Building the frontier ecosystemblogs.microsoft.com
- AnthropicPrimaryAnthropic prices initial public offeringanthropic.com
- AnthropicPrimaryAnthropic announces closing of initial public offering and concurrent private placementanthropic.com
- NVIDIAPrimaryNVIDIA Announces Financial Results for Third Quarter Fiscal 2027nvidianews.nvidia.com
- NVIDIAPrimaryNVIDIA Board of Directors Establishes National Security and Export Control Oversight Committeenvidianews.nvidia.com
- The DecoderOpenAI launches GPT-6.2 Astra and Sol with lower API prices and some benchmark gainsthe-decoder.com
- The DecoderGoogle ships Gemini 4 Neon, its fourth new Gemini model in a monththe-decoder.com
- TechCrunchAnthropic prices IPO at $83 a share, above projected rangetechcrunch.com
- CNBCNvidia closes lower ahead of earnings as company reels from Huang resignationcnbc.com
- weekly-recap
- model-pricing
- agents
- ipo
- earnings
- OpenAI
- Microsoft
- Hugging Face
- OpenRouter
- Anthropic
- NVIDIA
- GPT-6.2 Astra
- GPT-6.2 Sol
- Gemini 4 Neon
- Code 2
- Code 2 Flash
- Microsoft-Decision-1.1