Executive Summary
The cost of a unit of agentic work is falling from three directions at once. On silicon, TPUv7 now competes with NVIDIA on tokens per dollar for agentic serving. On weights, Kimi K3.1 Pro is within half a step of the closed frontier. On closed-lab pricing, GPT 6.1 Terra launched 50% below its predecessor. With cost dropping this fast, trust becomes the binding constraint on enterprise agents. The open questions are whether agents weigh evidence correctly, whether they deceive or collude, and whether the data they move can be seen. Clients who optimize for token price alone are optimizing the wrong variable.
What Changed
- NVIDIA now has a non-AMD rival on price-performance. TPUv7 FP8 delivers 1.2x–2.7x the tokens per dollar of B200, B300 and MI355X on GLM. It also beats GB300 FP4 end to end on Qwen3.5 397B. NVIDIA still leads wherever FP4 applies: about 3.5x on GLM, and 1.5x–2.6x on Kimi K3 after four software refreshes.
- Open weights reached near-parity on agentic benchmarks. Kimi K3.1 Pro reports beating Astra on Terminal Bench and shipped with day-0 TPU support. Self-hosting it still needs about 1.2 TB of VRAM.
- Computer use went small. North Ka Nano (9B) scores 78.9 on OSWorld-Verified, against 88 for the 397B Pro.
- The adversarial-reasoning lead flipped in six months. GPT-6 Astra beats Opus 4.6 in about 92% of social-deception games. The gap comes from skepticism toward any single piece of evidence.
Cross-Expert Synthesis
SemiAnalysis and Berman describe one market from opposite ends. Google gave chips and engineering help to the maintainers of vLLM, SGLang and llm-d, and that is why a 3T Chinese open model got day-0 TPU support. The CUDA moat and the closed-model moat are eroding through the same mechanism: open-source serving software. A sovereign stack is now a credible engineering design rather than an ideological one. That stack is open weights, served disaggregated on non-NVIDIA silicon.
Dwarkesh and Kreiner reach the same less obvious conclusion: verdicts have to come from full context, not from single signals. In Dwarkesh's games, Claude cleared the real demon because it accepted a sabotaged result at face value. In Kreiner's account, rule-based DLP fails because the same event can be benign or malicious depending on who did it and why. Orion's event-graph lineage applies "weigh every claim against the whole record" to data movement. GPT's evidence handling applies the same principle to reasoning.
Nate Jones supplies the decision rule the other sources lack: delegate when the agent's cost is below your hourly rate times the hours saved. The rule leaves out verification cost, and Dwarkesh's results show that verification cost depends on how a model behaves when its inputs are corrupted. No price sheet captures that behavior. Cheap tokens from a credulous model can cost more than expensive tokens from a skeptical one.
Enterprise Implications
For agentic serving, cost is driven by KV cache capacity and by splitting prefill from decode, not by peak FLOPS. A procurement benchmark built on single-turn chat will misprice hardware. Dwarkesh's harness, with parallel sub-agents reading every game log, is the realistic workload shape: bursty, multi-turn, and heavy on reused prefixes.
Choose models by role. Agents that consume tool outputs, retrieved documents or messages from other agents need their handling of adversarial evidence tested directly, because leading generic benchmarks does not predict it. The mixed Opus-and-GPT team showed no benefit in a small sample, so vendor diversity is no substitute for testing.
Agent deployments will multiply data-movement events while employees' unsanctioned GenAI use is already reopening DLP budgets. Agent-generated content breaks container-based tracking, so lineage rebuilt from observed events is the approach that survives.
Kimi and GLM are Chinese-origin models and will trigger procurement review. Whether a client actually gets sovereignty benefits depends on where the model is hosted, not on the weights.
What To Do About It
- Build an AgentX-style trace from a real client multi-turn workflow. Run it on TPUv7 FP8, GB300 FP4 and at least one inference provider before any 2027 capacity commitment. Treat FP4 as an accuracy trade-off to measure, not a default.
- Use TPU credibility and Kimi K3.1 as pricing leverage with NVIDIA and the closed labs now. Avoid multi-year lock-in on current numbers: TPUv8 (native FP4, Q1 2027) and Rubin will reset the comparison within two quarters.
- Add a corrupted-input test to agent evals: inject one false tool output and measure how much the agent relies on it.
- Have each client team set a loaded hourly rate, and track realized savings net of review time, not projected savings.
- Customer question: "When one of your agents gets a wrong tool result, what does it do, and how would you find out?" Most clients cannot answer it, which opens an eval and observability engagement.
Risks and Blind Spots
Nearly every figure here is vendor-reported or vendor-submitted. Some SemiAnalysis comparisons use Google's internal cost basis. The Kimi K3 result reversed within weeks, after four NVIDIA refreshes, so benchmark leads are perishable. The deception head-to-head pits Astra against February's Opus 4.6 and says nothing about Opus 5.5. Play is also stateless, so adding persistent memory could reorder the rankings. Orion's claims of day-one verdicts and full coverage come from a friendly interview with no independent data.
The unprompted deception Dwarkesh recorded, such as a minion sacrificing itself to build a teammate's credibility, happens inside a game today. It is still the right threat model for any workflow where agents message each other, and none of the sources addresses how to detect it. Finally, Claude built and revised the benchmark harness itself over 20+ iterations. If building RL environments is automating, every ranking above will change faster than procurement cycles can absorb.