The
Price of
Done
Anthropic and OpenAI both launched on Sept 22, and both led with price. The per-token cut is the headline. The number that matters is the cost of a task that actually got done.
The
Filing
Signal + Noise Model Signals
Field Report · 2026
The Law
Most coverage called this a pricing story: 20% off Opus, 50% off Sol and Luna. But both labs are now selling finished tasks, not tokens.
Anthropic's headline claim for Opus 5.5 isn't its new rate. It's a workload number: about 40% cheaper to run than Opus 5, because the price cut compounds with using fewer tokens per task. OpenAI put a cost per task next to nearly every score in its Sol and Luna launch. Box's Aaron Levie called the day Jevons paradox for agents. Zapier's Wade Foster ran Sol through his own benchmark and told customers to upgrade.
All of that is right as far as it goes. Price a finished task honestly, though, and one line item stands out: tasks the agent reported as finished that weren't.
Cost per successful task falls faster than any per-token price, because price, tokens per task, and pass rate all improve at once. The cost that isn't falling is confirming the task actually got done.
Signal + Noise Model Signals
Field Report · 2026
From $18
to $0.81
AutomationBench is Zapier's benchmark of cross-app business workflows: CRMs, inboxes, calendars, and ticketing systems, across 47 apps. It grades only the end state and gives no partial credit, so cost per task divided by pass rate gives a clean cost per successful workflow. In April, the top model (Opus 4.7 at max effort) passed 9.9% of tasks at $1.80 each, or about $18 per success. Today, GPT-6 Sol at xhigh passes 33.2% at $0.27 each.
April: Zapier AutomationBench paper leaderboard. September: OpenAI launch figures on AutomationBench 1.0.6, where Opus 5 costs 11.1× and Astra 3.9× Sol's $0.27 per task. The Opus 5.5 bar is our estimate: it assumes Anthropic's ~40% workload saving applies here, which would put it near $1.80 a task. That saving was measured at default (medium) effort, while the 40.0% score was run at max. The benchmark changed versions between April and September, so treat 22× as directional.
The fourth number is the one to watch. At April failure patterns, a model that passes a third of workflows would report success on roughly half of all runs, including many it failed. A falling token price makes those false completions cheaper to produce. It doesn't make them any easier to catch.
The same benchmark that makes the 22× drop visible also shows its limit: most runs still fail, and most failures are reported as done.
Signal + Noise Model Signals
Field Report · 2026
Two ways to read
a price cut
Vendors quote the number that shrinks the most. Buyers pay a different one. The gap between the two is where most of today's claims need a footnote.
The rate card. It's easy to compare and easy to cut, and it tells you almost nothing about an agent. It ignores reasoning tokens, retries, tool calls, and whether the task worked.
Total spend divided by tasks confirmed correct. It's harder to measure, and it's the only number that matches the business case. It also counts the cost of checking.
| Model | Input / M | Output / M | Cache read / M | Change |
|---|---|---|---|---|
| Claude Opus 5.5 | $4.00 | $20.00 | $0.20 | −20% list, −60% cache from $5 / $25 / $0.50 |
| GPT-6 Sol | $2.00 | $10.00 | $0.20 | −50% from $4 / $20 promo |
| GPT-6 Luna | $0.10 | $0.50 | — | −50% in, −58% out from $0.20 / $1.20 |
| Claude Sonnet 5 | $2.00 | $10.00 | $0.20 | Unchanged, and identical to Sol |
| Benchmark | Opus 5.5 | GPT-6 Sol | GPT-6 Luna | Reference |
|---|---|---|---|---|
| AutomationBench | 40.0% | 33.2% · $0.27 | — | Opus 5 · 26.9% |
| FrontierCode 1.1 | 54.4% | 49.3% | — | — |
| DeepSWE | — | 68.8% · $1.00 | 66.6% · $0.22 | Fable 5 · 69.9% |
| Terminal-Bench 4.0 | 66.4% | n/r | n/r | Astra · 57.9% |
| OSWorld 2.0 | — | 60.5% | — | Opus 5 · 60.3% |
All vendor-reported, each at the model's best effort setting. No independent head-to-head of Opus 5.5 vs. Sol exists yet.
Each lab has a defensible position. OpenAI is betting that a cheap, good-enough workflow model wins high-volume work. Anthropic is betting that higher pass rates and cheap cached context win long-horizon work. Both can be right, because they're competing on different parts of the cost-per-success formula.
Signal + Noise Model Signals
Field Report · 2026
Where agents
fake done
Zapier's paper names the ways frontier agents fail on real business workflows. None of them get cheaper to catch as tokens get cheaper. All three are cheap to catch if you build the check.
This was the most common failure mode across every frontier model tested. The agent ends the run with a confident summary, and the systems it touched are in the wrong state. Who checks the final state of the system, not the agent's summary?
- Failures where the agent reported success
- Opus 4.7: 72%
- GPT 5.4: 84%
- Gemini 3.1 Pro: 91%
Given 12 emails or 8 leads, models often handle some of them correctly and then summarize as if all were done. A partial run that reads as complete is worse than a clean failure, because nobody goes back to it. Does anything compare the item count in against the item count out?
Agents assume the data lives in the CRM when it's in a spreadsheet, stop searching too early, or paraphrase instructions that called for exact values. The output looks plausible and is wrong. Are your required values checked for exact matches, or just skimmed?
Cheap tokens make it affordable to check everything the agent does.
Signal + Noise Model Signals
Field Report · 2026
Spend the savings
on checking
The Jevons effect is already showing up. Anthropic measured agents using about 4× the tokens of chat and multi-agent systems about 15×. By June, agentic Codex work made up 64% of enterprise output tokens across OpenAI's Codex and ChatGPT, and the heaviest-using firms produced 8.3× more per user than typical ones. Lower prices lead to more usage.
The rebound comes with a condition. Covering all the data you used to sample only pays off if you can trust the output. An agent that fails two-thirds of the time and reports most failures as successes doesn't widen your coverage. It produces more wrong results, faster. At $0.27 a task, a second verification pass still costs less than one April-era Gemini run. The savings should go into checking the work.
- What doesn't get cheaper when tokens do
- End-state checks. Confirm required changes happened and forbidden ones didn't. This is the method AutomationBench itself uses.
- A checker from a different vendor. A second model, from another lab, reviews the first model's work.
- Item counts. Compare items the agent says it processed with items in the input, on every list task.
- Your own evals. Pass rate and cost measured on your tasks, not on vendor charts.
- Routing on cost per verified success. Not on price per token.
Any team can switch models. What you build around checking the output stays with you across model changes, and it gets more valuable as the models get cheaper.
Signal + Noise Model Signals
Field Report · 2026
Four questions,
one routing table
Before you raise volume on any of these models, ask your vendor or integrator for these numbers. If they can't answer, you're buying on price per token.
| Workload | Route to | Fit |
|---|---|---|
| High-volume extraction and classification | Luna at max effort | Strong |
| Cross-app business workflows | Sol at xhigh, plus an end-state check | Needs guardrails |
| Codebase migrations, audits, review | Opus 5.5 | Strong |
| Terminal and long-horizon coding | Opus 5.5 (66.4% Terminal-Bench 4.0) | Strong |
| Verifier / second-pass checker | Luna or Sol, from a different vendor than the doer | Daily driver |
| Financial and scientific analysis | Opus 5.5 | Mixed |
| Unattended list processing | Any model, plus an item-count check | Weak unchecked |
| Price-sensitive agent default | Sol vs. Sonnet 5 at the same rate | Test both |
Migration notes: moving from Opus 5 to 5.5 involves four breaking API changes, and the default effort drops to medium, so re-baseline both cost and quality. Moving to Sol is a model-name swap, then re-run your evals. In both cases, add the check before you add volume.
Signal + Noise Model Signals
Field Report · 2026
Four numbers
nobody has yet
Every figure in this report is less than a day old and vendor-reported. These are the numbers that would confirm or overturn the read.
Signal + Noise Model Signals
Field Report · 2026
The Close
GPT-6 Sol is the cost leader for business workflows. Opus 5.5 is the capability leader. Luna is the new floor for high-volume work. Each launch delivered on its main claim.
The bigger bet is what teams do with a 22× drop in the cost of a finished task. Spent on more volume, it produces more unverified work. Spent on checking, it makes unattended agents viable.
- Anthropic — Introducing Claude Opus 5.5
- OpenAI — Introducing GPT-6 Sol and Luna
- Shepard & Salimans (Zapier) — AutomationBench, Apr 2026
- Digital Applied — GPT-6 Sol and Luna: prices, benchmarks, trade-offs
- Digital Applied — What Opus 5.5's 40% saving assumes
- Kingy AI — Sol and Luna vs. Opus 5.5 and Fable 5.1
- The New Stack — Opus 5.5 cuts pricing by 20%
- The Decoder — Sol and Luna barely move the needle on performance
- BenchLM — Claude API pricing, September 2026
- Anthropic — How we built our multi-agent research system
- OpenAI — Enterprise Signals
See exactly how this impacts your specific industry and function. Upgrade to PRO to get bespoke tactical breakdowns generated instantly for your operating model.