Pick an agent model this week and every pricing page you open is quietly misleading you. Here's the rule that isn't: compare models on cost-per-completed-run, not price-per-token — because an agent's bill is tokens times turns, and the cheapest per-token model can be the most expensive per finished task.

That matters right now because four vendors are undercutting each other on the same seven days. As of the week of July 23, 2026: GPT-5.6 Sol is priced around a quarter of the prior flagship; Meta's Muse Spark 1.1 opened the Meta Model API at $1.25 input / $4.25 output per million tokens; Sonnet 5 is on a promotional $2/$10 through August 31; and Claude Opus 5 is rumored this week, positioned — if the reporting holds — on lowering the cost of long agent runs. The prices are all falling. The right way to choose did not change.

Why the pricing page lies#

A pricing page shows you dollars per million tokens. An agent doesn't spend tokens — it spends turns, each of which spends tokens. Total cost is roughly:

run cost ≈ Σ (input_tokens × in_rate + output_tokens × out_rate)  over every turn
true cost per outcome ≈ run cost ÷ success_rate

Two things live inside that sum that the pricing page can't show you:

  1. Turn count. A weaker-but-cheaper model loops more — it retries a failed tool call, re-reads context it dropped, takes eight steps where a stronger model takes four. Halve the per-token price, double the turns, and you've saved nothing.
  2. Failure rate. A run that ends in a wrong answer still cost money. If model A finishes 95% of tasks and model B finishes 70%, B's real cost per good outcome is far higher than its per-token discount suggests.

The cheapest model on the pricing page and the cheapest model in production are frequently not the same model. The gap is turns and retries.

Output tokens are where this bites hardest. Agents are output-heavy and context-heavy — they generate long tool calls and reasoning, and they carry big contexts. A model that's cheap on input and expensive on output can lose to a rival with a higher headline rate. Read the output column first.

The measurement, in one afternoon#

You don't need a benchmark harness. You need one representative task and a log.

  1. Pick one task that looks like your real work — the agent job you run most, with a clear success/failure check. Not a toy prompt.
  2. Run it end to end on each candidate, ideally 5–10 times to average out variance. Keep the scaffold, tools, and prompt identical; change only the model.
  3. Log four numbers per run: total input tokens, total output tokens, total turns, and wall-clock time. Price the tokens at each model's rates and sum to a dollar cost per run.
  4. Divide by success rate. Ten runs, eight correct → divide your average run cost by 0.8. That's your true cost per good outcome.
  5. Rank on that number, then sanity-check latency. For an interactive agent a user waits on, a slightly pricier model that finishes in half the wall-clock time is often the right call — Sol's throughput pitch is exactly this trade.

Most teams have never computed step 4. It routinely reorders the leaderboard: the model that looked 20% cheaper per token turns out 15% more expensive per completed task because it retried twice as often.

The picks, honestly#

The price war is good news — your agent bill is going to fall no matter which way you jump. Just don't let a lower sticker pick your model for you. Measure the run, not the token.

Related: the launch driving this week is covered in Claude Opus 5 is days away — and the pitch is your agent bill.