The Compute Bill Is Rearranging the AI Leaderboard
Three kinds of cost behind one leaderboard
The interesting part of the latest MathArena result is not who finished first. Google's Gemini 4 Carbon took the crown, with Anthropic's Fable 5.2 close behind, knocking the long-dominant GPT-6.1 Sol off the top. But the same reporting adds two details: both are significantly more expensive, with Gemini 4 Carbon costing 15 times more than GPT-6.1 Sol, and Gemini ignored the time limit during testing, almost never giving up on a task.
Those details carry more signal than the ranking itself — the top spot was bought with money and with a refusal to stop reasoning. For teams making thousands of calls a day, a 15x price gap outweighs any benchmark lead.
Training gets cheaper, using gets dearer
At the training end, the curve appears to bend the other way. Former OpenAI and Anthropic researcher Jacob Coxon says Anthropic can already retrain Mythos with "a fraction of the compute originally needed." He traces this to a slide from the company's IPO roadshow and cites it as his reason for leaving: the algorithms are not just improving quickly, but at an accelerating rate. He continues to call for a pause on frontier model development.
This deserves restraint: the "accelerating" claim is Coxon's judgment, not an independently verified fact. But if training costs really are falling while the cost of using top models is pushed up by reasoning length, a scissors gap opens — capability becomes cheaper to produce, and steadily using it at scale becomes more expensive.
Organizations shrink, capability concentrates
OpenAI dissolved its standalone robotics team for the second time. The stated reason is not a lack of data but that general-purpose models such as GPT-6 Astra can control robots directly, making a dedicated unit unnecessary; the researchers were redistributed across other teams.
Anthropic, meanwhile, made Claude Opus 5.5 the default model on all paid plans and published a guide telling users to drop old habits: stop writing "think hard" and start at Medium effort.
Both point the same way: vendors want capability to move from something requiring a dedicated team and dedicated prompting tricks to something that is simply there by default. The direct consequence for users is that previously accumulated prompting habits may need rewriting.
Cracks on the capital side
Days before what could be the largest IPO in history, Amodei admitted "there's a huge uncertainty in what's going to happen next year," and called buying compute "the deepest business dilemma" his company faces. Anthropic's commitments now total nearly $901 billion, and it remains unclear whether future revenue will justify those costs.
On the Nvidia side, Jensen Huang's departure is being read by VCs and founders as a funding hole: under his leadership Nvidia poured tens of billions into AI startups, and founders now brace for smaller checks and lower valuations. Asian governments have reacted unevenly — some moved to confirm chip supply commitments, others are assessing the impact on local projects or addressing ongoing investigations.
Conclusions and action items
- Put cost into model selection. A high-score, high-price combination like the MathArena leader is a benchmark, not a default. Run a unit-economics calculation with a concrete number like the 15x gap before choosing a primary model.
- Rewrite your prompts. The default model is now Claude Opus 5.5; starting at Medium effort and removing instructions like "think hard" is a zero-cost migration.
- Reassess dedicated vertical teams. Signals like general-purpose models taking over robot control mean the marginal value of a bespoke vertical AI team is falling, while single-vendor dependency risk is rising. Price both together.
- Assume a more conservative funding environment. LangChain's $3B valuation and 7,000 LangSmith customers show the agent-management layer still holds independent value, but the broader private market may tighten — extend your runway math.
- Use the graveyard as a reference frame. The repeated delays to Siri AI and the messy launch of OpenAI's "super app" both suggest that shipping cadence is not the same as capability. Evaluate vendors on delivery, not announcements.
Sources
- Gemini 4 Carbon takes the MathArena crown but costs 15 times more than GPT-6.1 Sol
- Anthropic can already retrain Mythos with a fraction of the compute originally needed, says former researcher Coxon
- Anthropic says outdated prompting habits are slowing down Claude Opus 5.5
- OpenAI reassigns robotics researchers to other teams as general-purpose models might make specialized robotics AI unnecessary
- Anthropic CEO Amodei admits there's a huge uncertainty in what's going to happen next year days before a record IPO
- AI agent startup LangChain hits $3B valuation
Subscribe to weekly briefing