The Agent Price War: What a $2.49 Task Means for Indie Builders
The most interesting number in AI right now isn't a benchmark score. It's $2.49.
That's the reported cost of a completed agentic task on Grok 4.5. The same task on Claude Fable 5 — the model I called a ceiling-mover just six weeks ago — reportedly runs about $11.80. One developer shared that a batch of tasks costing him $63 on Fable came in at $16 on GPT-5.6 Sol.
Take the exact figures with a grain of salt — they're community-reported, task-dependent, and every vendor measures "completed" differently. But the direction is unmistakable, and the labs are behaving like it's true. That's the tell.
Anthropic blinked. Three times.
If you followed the Fable 5 saga this month, you watched a price war play out in real time, disguised as a support-page deadline.
Quick recap: Fable 5 launched June 9, went dark three days later under export controls, and came back July 1 with included subscription access "through July 7." Then July 7 became July 12. Then July 12 became July 19. Three deadlines, three last-minute extensions — each one landing suspiciously close to an OpenAI announcement.
On July 18 it finally resolved: Max and Team Premium plans keep Fable 5 permanently at 50% of weekly limits, Pro moves to usage credits with a one-time $100 cushion. A real decision, at last.
I don't think Anthropic was being indecisive for fun. Read the moves together — a reported $45B SpaceX compute deal, leaked talks to lease up to $10B of capacity from Meta, usage limits as the public face of a compute crunch — and the picture is a company squeezed between the best model on the market and competitors who can serve tasks at a fifth of the price. When your rival's flagship costs $2.49 a task, "our model is better" has to be a lot better.
The metric quietly changed
Here's what I think actually matters, beyond the corporate drama.
For three years we priced AI in tokens. Tokens were always a proxy — nobody wakes up wanting a million tokens. You want the bug fixed, the report written, the pipeline green.
The industry has now converged on the metric that was always the real one: cost per completed task. That's what the $2.49 vs $11.80 comparison measures. Not throughput. Not context window. Did the agent finish the job, and what did it cost end to end — including the retries, the tool calls, the wandering.
That framing punishes different things than token pricing does. A cheap model that fails 40% of the time and burns three attempts isn't cheap. An expensive model that one-shots a task might be the budget option. The unit economics of agents are about reliability per dollar, and that's a much harder thing to fake on a leaderboard.
This is the same shift I keep seeing everywhere this year: the conversation moved from "how big is the model" to "how well does it complete real work without supervision." The price war is that shift showing up on an invoice.
Why indie builders are the accidental winners
Big enterprises negotiate contracts. We don't. We pay list price and live inside usage limits. Which means when frontier labs start undercutting each other on task pricing, we're the ones who feel it first — in a good way.
Three practical consequences:
1. The frontier war is subsidizing your compute. Use it — but don't build on it.
Free Fable access extended three times. Weekly limits reset "to coincide with" a competitor's launch. A $100 credit handed to every Pro user. This is promotional pricing in a land-grab phase, the same movie we watched with cloud credits and ride-sharing. Absolutely spend the subsidy — prototype on the frontier models while they're artificially cheap. But if your product's unit economics only work at promo pricing, you don't have unit economics. Model your costs at list price and treat everything below that as margin, not a plan.
2. Route by task, not by loyalty.
The most quietly instructive release this month was Microsoft's Project Perception — a security tool that mixes Anthropic, OpenAI, and Microsoft's own models, routing each step to whatever is cheapest for that step. Microsoft using Anthropic's models to undercut Anthropic's own product is peak 2026, but the pattern is the point: smart routing is what makes continuous agentic workloads affordable. Run the cheap model for the 80% of steps that are routine, escalate to the expensive one when it fails or when the step actually demands it. For my own agent work, that alone is the difference between "interesting demo" and "runs every night."
3. Measure your own cost-per-completed-task. Today.
The vendors' numbers are marketing. Yours aren't. If you're running agents in production and can't tell me what a completed task costs you on each model — including failed attempts — you can't make routing decisions, and you can't tell whether the next price cut changes anything for you. It's an afternoon of logging. Log the model, the outcome, the retries, and the total spend per task. That table is worth more than any leaderboard.
And a fourth, structural one: build vendor-agnostic from day one. After watching the best model on the market disappear for nineteen days over export controls, then live behind a rolling deadline for three weeks — being hard-wired to a single provider isn't a simplification, it's a liability. An abstraction layer over your model calls costs you a day now and saves you a rewrite later.
The honest caveats
Because I'd rather you trust the argument than the numbers:
The $2.49 and $11.80 figures come from community reporting, not audited benchmarks. Task mix matters enormously — Fable 5 still wins on the long, complex, autonomous work it was built for, and for those tasks "expensive per task" can still be the cheapest path to done. Cheapest-per-task also isn't cheapest-per-correct-task; a model that quietly ships wrong answers is infinitely expensive. And promo pricing distorts every comparison made during a land grab.
None of that changes the conclusion. It sharpens it: the only numbers that matter are the ones from your workload.
Where this goes
Inference costs have been falling off a cliff for two years, and the price war is just that curve reaching the frontier tier. My bet: within a year, cost-per-completed-task becomes the number on the pricing page, routing becomes a default feature of every agent framework rather than something you hand-roll, and the labs compete on reliability curves the way chip makers compete on yield.
For indie builders, that's the best possible outcome. The expensive part of the last three years wasn't the models — it was that serious agentic workloads were priced for enterprises. That's ending. The work I couldn't afford to automate in January is affordable in July.
The ceiling moved again. This time it was the price.
Running agents in production and tracking your cost-per-task? I'd genuinely like to see your numbers — tell me where my math falls short.