The Cheapest Frontier Model in August 2026: 11 Configurations, One Battery, 17x Cost Spread

Over the past week we've run eleven model configurations through an identical battery of three hard, script-graded tasks. Same prompts, same max_tokens, same grading code, no LLM judge.

All eleven passed all three tasks. Here is what they cost.

The table

17.1× between cheapest and most expensive, for outcomes our grader cannot tell apart.

What the ranking is not

It is not a quality ranking, because there was no quality difference to rank. Every configuration found the interval-merge bug that only fires on a fully-contained interval, counted the letter "r" correctly in a phrase where the answer is 10, and produced four sentences of exactly five words each without using a forbidden word.

It is also not a claim that these models are interchangeable in general. Our battery is single-turn and resolves in seconds. It cannot measure long-horizon agentic work, tool-use chains, sustained coding sessions, or anything where small per-step differences compound — which is exactly where the expensive models are sold, and exactly what their published benchmark leads measure.

What it does establish: for ordinary, verifiable, one-shot production work, the frontier has converged, and price is no longer a proxy for capability.

Three things the price list won't tell you

1. The rate card is a poor predictor of the bill

Sort the table by list output rate and you get a different order than sorting by measured cost. Some examples from the runs behind it:

• Kimi K3 charges $15/M output. GPT-5.6 Sol charges $30/M. K3 cost more, because it emitted 3.1× the tokens.
• Grok 4.5's list rate is 4.2× below Opus 5's. Measured, it came in 11.4× cheaper — it's unusually terse.
• GPT-5.6 Luna lists 5× below Sol. Measured on an easier battery, that gap was 3.3×, because Luna is verbose.
• Gemini 3.6 Flash cut its output rate 16.7% versus 3.5 Flash. The measured saving was 7.7%.

Verbosity is the hidden multiplier. A model's token appetite varies by a factor of six across this table — 286 to 1,751 output tokens for the same three answers — and you pay for every one.

2. The same model's cost moves between runs

We measured Claude Opus 5 four separate times on this identical battery: $0.035025, $0.037125, $0.037225, $0.044200. A 26% spread, same model, same prompts.

If you're comparing two models and the gap between them is under about 25%, one run cannot tell them apart. Most published comparisons — including some of ours — report a single run.

3. Where you rent it can matter more than what you rent

Kimi K3 costs the same on Together and Fireworks. Fireworks served it exactly 2.00× faster. Identical weights, identical price, double the wait. That variable appears on no leaderboard.

Reading the table for your own workload

High-volume, latency-tolerant — Grok 4.5, Gemini Flash, or GPT-5.6 Luna, all clustered around $0.0033–$0.0039. At a million calls shaped like this battery that's roughly $3,300, against $56,100 for Fable 5.

Interactive paths — cost is not your constraint, latency is, and the two don't correlate here. GPT-5.6 Terra and Luna were the quickest in our runs at 8.1–8.8 seconds for the battery; Kimi K3 on Together took 69.8 seconds for the same work.

Hard output rules — pick on measured constraint adherence, not price. Both Gemini Flash models went 10/10 on a constraint test where Opus 5 has scored 13 of 29 across five runs. The most expensive model in the table is the least reliable one on that axis.

Genuinely frontier problems — this table has nothing useful to say. If your work is long-horizon and agentic, trust the vendor benchmarks over our battery, because they're measuring something ours can't reach.

One date to note: Sonnet 5's $2/$10 is introductory and ends 2026-08-31. At its standard $3/$15 its measured cost here becomes roughly $0.037 — level with Opus 5's cheapest run.

Run it yourself

Every prompt is reproducible with no account and no API key:

• The interval-bug task
• The letter-counting task
• The constraint task

To run this comparison across models on your own prompts and see the token counts that drive the bill, that's Promptster.

All figures measured 2026-08-07 to 2026-08-09 via the Promptster API, max_tokens: 4000, on an identical three-task battery. Costs are computed from list pricing at time of writing and reflect actual token usage, not estimates. Ranges are shown where a configuration was measured more than once; single-run figures should be read with the 26% run-to-run variance noted above in mind. Latency figures referenced in the text come from different sessions and are indicative rather than a controlled comparison — cost is deterministic from token counts, latency is not. Sonnet 5 is priced at its introductory rate, which ends 2026-08-31.