The AI 'value war': measure cost per task, not per token
OpenAI, Meta and SpaceX are racing to be cheaper per unit of work than Anthropic. Bloomberg calls it a value war — here's how to actually benchmark your bill.
The last week of model launches had a strange headline feature: the pitch wasn't raw capability, it was how little the model charges to do the work. Bloomberg framed it bluntly on July 12 — OpenAI, Meta, and SpaceX are competing to build the most cost-efficient AI, not the smartest. For anyone paying an AI bill, that shift matters more than any benchmark score. The frontier is now competing on cost per task, and that's the number you should be measuring too.
What actually happened
OpenAI's flagship GPT-5.6 Sol is built to finish more work using fewer tokens. On the Terminal-Bench 2.1 suite it edged Anthropic's Opus 5 on accuracy (roughly 88.8% to 88%) while solving the full set for a total token cost of about half — the efficiency, not the sticker price, is the weapon. Meta's Muse Spark 1.1 landed at roughly a quarter of flagship rivals' pricing, and SpaceX's Grok 4.5 at $2/$6 per million tokens. Bloomberg describes the market reshuffling into a "two strong, three chasing" shape: Anthropic and OpenAI leading on capability, with Google, Meta, and SpaceX pushing lower-cost models underneath them.
The demand side is driving it. As CNBC has reported, business customers are scrutinizing AI spend and shifting toward efficiency — so the vendors are optimizing for tokens-per-task instead of chasing another point of benchmark accuracy.
Why it matters for your business
Per-token price is a trap. A model that costs more per million tokens but finishes your task in a third of the tokens is cheaper per outcome — and the reverse is just as common. The only figure that tells you the truth is what it costs to complete one unit of your actual work: a resolved support ticket, a generated product description, a fixed bug. Benchmark that on your own workload, not the vendor's demo.
There's an upside hiding here. Because the leaders are now competing on efficiency, your current vendor may be getting cheaper per task without you switching anything — but you'll only notice if you're tracking cost-per-outcome over time. And the way to actually shop this price war is the move we always come back to: keep every model behind an abstraction you own, so testing GPT-5.6 Sol against Opus 5 against Muse Spark on your jobs is a config change and a spreadsheet, not a migration.
Key takeaways
- Bloomberg (July 12): OpenAI, Meta, and SpaceX are competing on cost-efficiency, not raw capability — a "value war"
- GPT-5.6 Sol matched Opus 5's accuracy on Terminal-Bench 2.1 at roughly half the total token cost; Muse Spark 1.1 and Grok 4.5 undercut on price too
- Per-token price lies — a model can cost more per token and less per finished task. Measure cost-per-outcome on your own work
- Keep models behind an abstraction you own so you can A/B them on your jobs and capture the price war as savings
Do you know what one unit of your AI work actually costs? We build automation behind a model layer you own — instrumented for cost-per-task, so you can route to the cheapest capable model and prove the savings. Estimate what owning your model layer is worth or see how we build it.
- #ai-pricing
- #model-routing
- #cost-optimization
- #gpt-5-6
- #anthropic
Tommy Rush — Founder, Rush Commerce
Operator turned builder. 15+ years running operations — now shipping the systems businesses run on. More
Get The Rush Report weekly — one email, zero fluff.
Keep reading
Quiq ships an AI agent control layer. Govern before go-live.
Quiq's Verified Intelligence adds guardrails, conversation simulations, and full decision logging to customer-facing AI agents. Why the control layer is the real product.
Read itAuger's $50M bet: unify your systems, don't rip them out
Ex-Amazon ops chief's startup Auger raised $50M to sit on top of ERP, WMS and TMS and orchestrate them — the integrate-don't-replace thesis, now funded.
Read it