AI Frontier

What's new in AI, and why it matters.

Updated Nov 16, 2026, 8:35 AM PT

Dark hero graphic. A three-step date strip: Tuesday Nov 17, order book expected to close; Wednesday Nov 18, pricing expected in a $70 to $78 a share range; Thursday Nov 19, highlighted in red, first trade on Nasdaq under ANT. Below, a red step line labeled reported, unnamed sources shows orders as a multiple of shares on offer: about 2 times on Nov 10, about 6 times on Nov 12 and about 7 times on Nov 15, with diamond markers for Nov 17, 18 and 19.

Today's lead

Anthropic sets its IPO for Nov. 19 at up to a $1.4 trillion value

The Claude maker will sell 825 million shares at $70 to $78 on Nasdaq as 'ANT', one of the biggest stock debuts ever.

Source: CNBC

Graphic: Plot · Data: CNBC; order book via The Next Web (citing Bloomberg and Reuters), reported, unnamed sources

Latest

Dark infographic titled “Frontier AI prices span 100×: $50 to $0.50 per million output tokens.” Paired horizontal bars show vendor list prices in US dollars per 1 million tokens (input, output) for 10 models, sorted by output price: GPT-6 Astra $10 / $50; Claude Fable 5.2 $8 / $40; GPT-6.1 Sol $2 / $10; Gemini 4 Carbon $2 / $10 through Nov 30, then $4 / $20; Gemini 4 Hydrogen $1.50 / $7.50; Grok 4.7 $2 / $6; Qwen4-Max $1.50 / $4.50; DeepSeek V4 Pro $1.32 / $3.96 at peak hours; Claude Haiku 5.5 $0.10 / $0.50; GPT-6 Luna $0.10 / $0.50. The two cheapest, Claude Haiku 5.5 and GPT-6 Luna, are highlighted in red. Callouts: highest output price $50, lowest $0.50, a 100× gap (our calculation).

Chart of the day

What frontier AI costs now: vendor API list prices for 10 current models from six labs range from $50 to $0.50 per million output tokens, a 100× spread (our calculation). Google's Gemini 4 Carbon doubles to $4 / $20 on Dec 1. List prices only; tier and region caveats apply.

Graphic: Plot · Data: vendor pricing pages (Anthropic, OpenAI, Google, Alibaba Cloud, xAI, DeepSeek) (self-reported), opened Nov 16, 2026

Model & benchmark comparison

Bar chart comparing four AI models on three tests run independently by Epoch AI. FrontierMath Tiers 1–3: Claude Fable 5.2 95.1%, GPT-6.1 Sol 93.7%, Gemini 4 Carbon 90.5%, Kimi K3.1 Pro 78.2%. GPQA Diamond: Fable 5.2 93.9%, GPT-6.1 Sol 95.4%, Gemini 4 Carbon 96.5%, Kimi K3.1 Pro 94.2%, against a PhD-expert score of 69.7%. SimpleQA Verified: Fable 5.2 94.0%, GPT-6.1 Sol 73.9%, Gemini 4 Carbon 77.7%, Kimi K3.1 Pro 62.2%.
On Epoch AI's own runs, the newest frontier models are near the ceiling on expert math and science questions. Factual recall (SimpleQA Verified) still separates them, with Claude Fable 5.2 well ahead.

Source: Epoch AI Benchmarking Hub (CC BY 4.0)

Timeline

What changed recently

  1. Sep 30, 2026

    Google unveils Gemini 4 Argon

    Google's new frontier model launched a new naming scheme and a 1M-token output limit.

  2. Nov 10, 2026

    Google releases Gemini 4 Carbon

    Google's most capable model so far; its API price doubles after Nov 30.

  3. Nov 12, 2026

    Jensen Huang resigns as Nvidia CEO

    CFO Colette Kress is interim CEO while the board searches for a successor.

  4. Nov 13, 2026

    ChatGPT Wallet lets agents pay

    OpenAI's agents can check out for you after you approve each purchase with Face ID or a fingerprint.

  5. Nov 19, 2026

    Anthropic IPO (upcoming)

    Shares are set to start trading on Nasdaq as 'ANT'.

Visual explainer

Large language models write by predicting one token at a time, then feeding each pick back in and predicting again. The probabilities shown are illustrative.

Four-step diagram titled 'How a chatbot writes: one token at a time'. Step 1 splits 'The cat sat on the' into tokens. Step 2 shows attention arcs from the last word back to earlier words. Step 3 shows example probabilities for the next token (mat 41%, floor 17%, sofa 9%, bed 6%, roof 3%), marked as illustrative. Step 4 adds 'mat' and loops back to repeat. A black strip notes that 1 token is about 4 characters, or three-quarters of an English word, and that fluent answers can still be wrong.

Source: OpenAI Help Center, 'Understanding and counting tokens'