Accel set to lead $5 billion into Thinking Machines at a $40 billion valuation

The venture firm Accel is set to lead a $5 billion round into Thinking Machines LabPVT$13.50B — Thinking Machines Lab, private, latest valuation $13.50B at a $40 billion valuation before the money arrives, Business Insider reported on November 6, 2026↗; the company, founded by former OpenAIPVT$1.18T — OpenAI, private, latest valuation $1.18T chief technology officer Mira Murati, declined to comment and Accel did not respond ¹².

The mark was set the night after Inkling 2 Small went live on Tinker, the lab's paid service for retraining models on customers' own data, with the weights, the files needed to run a copy, withheld pending outside safety testing ²³⁴. The wager is defensible, but it rests on the service rather than the benchmark headline, and the service bills by the token.

The cheap token buys an expensive answer

Inkling 2 Small uses a mixture-of-experts design, which routes each query to a few specialist sub-networks, so 17 billion of its 594 billion parameters work at a time ⁴⁵. Artificial Analysis, an independent model-benchmarking firm, scores it 31 on its intelligence index↗, well above the median of 24 for its price tier ⁶.

Per task, the bill is another matter: $1.18 of compute against $0.20 for the earlier Inkling-Small, because the new model writes about 150,000 output tokens a task where the old one wrote 37,000 ⁴⁶. On Terminal-Bench, a test of agentic coding in a live terminal, it scores 13 percent against Claude Sonnet 5.5's 30 ⁴. Tinker prices the model at $0.30 per million input tokens and $1.00 per million output ⁴⁶. Because the vendor sells tokens, a longer answer raises the customer's bill and the lab's revenue with it.

What the money is being asked to buy

The reported price is 20 percent below the $50 billion the lab sought, and abandoned, in late 2025, though well above its July 2025 seed; Accel's talks at that valuation floor first surfaced in September, when the round under discussion was $1 billion ²⁷. A September report put annualized revenue above $100 million↗ ²⁷. Against that, the valuation, not the check, is the multiple: roughly 400 times sales ²⁷.

Underneath the price sit an undisclosed Nvidia▼−4.65% — Nvidia, down 4.65 percent today investment from March, at least a gigawatt of Vera Rubin, the Nvidia platform due to follow its current Blackwell chips, from early 2027, and The Information's report that Nvidia weighed about $2.5 billion of this round ¹². Four of the lab's six named cofounders have left ².

The durable asset is Tinker: the hedge fund Bridgewater used it to fine-tune an open Qwen3 model for financial filtering, beating GPT-5.5's accuracy on six tasks at about a fourteenth of the inference cost ⁵. Among open models, though, the new Inkling placed second on four of six code benchmarks and first on none ⁵.

The download is the first real test

The lab's own page says↗ the model "is closed weight at release", with the full files due "in the coming weeks" under the permissive Apache 2.0 license, once outside safety testing ends ⁴⁸. The first Inkling shipped open on day one in July ².

This week also set the public comparison. AnthropicPVT$350B — Anthropic, private, latest valuation $350B, maker of Claude, filed on November 10 for an IPO↗ valuing it at up to $1.4 trillion on roughly $81 billion of run-rate revenue ⁹. That filing is the test for the private mark: about 400 times reported sales for Thinking Machines, against about 17 times run-rate revenue at the top of Anthropic's range ²⁹. The gap between those multiples is a growth claim, and nothing published yet substantiates it.

Fine-tuning as a business is real, and this round bets it compounds at gigawatt scale. The date to watch is the weights release: once the files are public, anyone with the hardware can run the model and judge it for themselves.