The frontier is the handful of laboratories whose models define what is possible this quarter and whose prices define what it costs. This briefing maps them at the end of 2025: what each shipped, what it charges, what it raised and what it has committed to build. It is a snapshot, and the year it describes began with a $6m model from Hangzhou that nobody in this list had planned for. Nomad is a member of Anthropic's Claude Builder Ambassador programme; the rankings below are the evaluators', and the advice at the end is model-agnostic because our business is.
The models
| Lab | Model | Released | Price in / out | Weights |
|---|---|---|---|---|
| Google DeepMind | Gemini 3 Pro | 18 Nov 2025 | Not compared here | Closed |
| OpenAI | GPT-5.2 | 11 Dec 2025 | Not compared here | Closed |
| OpenAI | GPT-5 | 7 Aug 2025 | $1.25 / $10 | Closed |
| OpenAI | gpt-oss-120b | 5 Aug 2025 | Hosted by third parties | Apache-2.0 |
| Anthropic | Claude Opus 4.5 | 24 Nov 2025 | $5 / $25 | Closed |
| Anthropic | Claude Sonnet 4.5 | 29 Sep 2025 | $3 / $15 | Closed |
| Anthropic | Claude Haiku 4.5 | 15 Oct 2025 | $1 / $5 | Closed |
| Meta | Llama 4 Maverick | 5 Apr 2025 | Hosted by third parties | Llama licence |
| Mistral | Mistral Large 3 | 2 Dec 2025 | Hosted by third parties | Apache-2.0 |
| Amazon | Nova 2 Pro | 2 Dec 2025 | See Bedrock pricing | Closed |
| DeepSeek | R1 / V3 family | Jan 2025 onward | $0.55 / $2.19 for R1 at launch | MIT |
The year's shape was a relay. OpenAI's GPT-5 in August introduced a router between fast and reasoning modes at a price a fraction of its predecessors', and its first open-weight models in six years the same week.[3][4] Anthropic's Claude 4 generation arrived in May and was refreshed through the autumn, ending with Opus 4.5 in November at a third of the previous flagship price and a reported score of about 80% on SWE-bench Verified.[5][11] Google's Gemini 3 Pro on 18 November reported 37.5% on Humanity's Last Exam without tools and 91.9% on GPQA Diamond, and took the top of the Arena.[1] OpenAI replied with GPT-5.1 and then GPT-5.2 on 11 December.[2] Mistral shipped a 675 billion parameter open model in December, Amazon a new Nova family the same day, and Microsoft announced its own superintelligence team in November.[9][10][12] The open-weight column, meanwhile, belongs increasingly to China, which is a briefing of its own.
The money
- $183bn
- Anthropic's valuation on its $13bn Series F in September, with run-rate revenue above $5bn [13]
- 27%
- Microsoft's stake in OpenAI Group after the 28 October conversion to a public-benefit corporation [14]
- $5tn
- Nvidia's market value on 29 October, the first company to reach it [15]
Anthropic's Series F on 2 September valued it at $183bn, three times its March valuation, on run-rate revenue that had gone from about $1bn at the start of the year to more than $5bn by August.[13] OpenAI completed the restructuring it had pursued all year on 28 October, becoming a public-benefit corporation under a non-profit parent with Microsoft holding about 27% and rights to its models through 2032.[14] The compute followed the money. OpenAI signed for some 33 gigawatts in three weeks of October, covered in the build-out briefing; Anthropic signed on 23 October for up to a million Google TPUs, with more than a gigawatt coming online in 2026.[16] Nvidia, which sells to all of them and invests in most, became the first $5 trillion company.[15]
Meta rebuilt itself
Meta began the year as the largest open-weight lab and ended it as something else. Llama 4 in April was received badly, and a leaderboard result that did not match the released weights became a case study in how not to launch.[7] In June the company paid about $14bn for 49% of Scale AI, made its founder Alexandr Wang chief AI officer and formed Meta Superintelligence Labs, with reported nine-figure offers to researchers. Behemoth, the two-trillion-parameter model previewed in April, never shipped. On 19 November Yann LeCun, Meta's chief AI scientist and a founder of the field, announced he was leaving to start his own company.[17] The next Meta model would be closed.
Who is buying what
The most useful single number of the year for a buyer came from Menlo Ventures on 9 December: enterprise spending on generative AI reached $37bn in 2025, three times the year before, and of the spend on model APIs Anthropic took 40%, OpenAI 27% and Google 21%, a reversal of 2023, driven by Anthropic's share of coding workloads.[18] Coding is where the models had clearly crossed into paid, daily use; agents were about a sixth of deployments. That story is in the year of the agent.
What a buyer should take from this
Four things. The gap at the top is small and changes monthly, so a decision made on this month's ranking will be wrong by the next; the tiers are where the real choice is, and on routine work the difference between a $5 model and a $1 one is usually invisible. The Chinese labs and Mistral offer the same capability class under open licences, hostable in the EU, at a fraction of the price. The labs are worth more than most countries' stock markets and are financing each other, which is a counterparty risk that belongs in a procurement note. And the benchmark that matters is the one run on your own documents.
