Moonshot cut Kimi's prices by a third. Artificial Analysis then measured the bill for an Intelligence Index task: $2.00 on K3, $1.98 on K3.1 Pro. A 33% discount produced a 1% saving — two cents — because output tokens per task rose from 48,000 to 55,000.
That is the whole story of this price war in one line. A rate card is an offer. A bill is a negotiation the model conducts on its own behalf, and it has been winning.
Last week we tracked who was cutting what as OpenAI and Anthropic pushed flagship API costs down as much as 80% and DeepSeek's V4.1 Flash spread to 29 inference providers on OpenRouter. This week the question changes: not who cut, but whether the cuts reach the bill at all. They mostly don't.
A rate card is an offer. A bill is a negotiation the model conducts on its own behalf, and it has been winning.
Start with Anthropic, because Anthropic disclosed the arithmetic itself. Per Artificial Analysis, Opus 5.5's roughly 80% increase in token usage would have pushed cost per Intelligence Index task to $10.51. A 20% base-price cut brought that to $8.41; cheaper cache reads brought it to $5.98. Opus 5 sat at $5.86. So a headline price cut and a cache-price cut together bought back almost exactly the verbosity the model added, and nothing more.
On agentic coding the mask slips entirely. On Artificial Analysis's Coding Agent Index, Opus 5.5 set a new high score of 66 and cost per task rose 21% to $13.04, driven by roughly 2.4 times more output tokens. The better model is the more expensive one to operate, and the price sheet says otherwise.
Bito found the same inversion from the buyer's side. Running Claude Fable 5.1 against Opus 5 across 60 engineering tasks, Fable's list price is double — $10/$50 per million tokens versus Opus's $5/$25 — and its bill came to $85.48 against $135.51. Both fully passed 49 of 60. The model with the higher rate card cost 37% less per finished task.
The mechanism is visible in usage logs. On a typical task Fable took 13 steps to Opus's 17, made 10.5 model calls to Opus's 20.8, and wrote about 6,700 output tokens against 21,600. Every step resends the conversation; a model that spirals pays for the spiral on every turn. Opus's 10 most expensive tasks accounted for 56% of its entire bill.
The pattern held at OpenAI. gpt-6-astra lists at $10/$50, gpt-5.6-sol at $4/$20. Sol's run cost $158.13 and scored 47.5; Astra's cost $129.26 and scored 54 — because Sol made 27 model calls per task to Astra's 17 and wrote four times the output. The model that looked 2.5x more expensive came out 30% cheaper per successful task.
Then there is DeepSeek, which solved this the hard way. V4.1 Flash scores 40 on Artificial Analysis's Intelligence Index, beating its own much larger V4 Pro, at $0.27 per index task. It is also one of the most verbose models measured — 89,000 tokens per task — and still cheap, because DeepSeek attacked token consumption rather than the rate card. Its own release claims token usage down more than 50% versus the prior version through task-aware compression, dynamic context trimming, and adaptive decoding.
Read the technical report and the theme is obsessive: KV cache compressed to 890 bytes per token globally, about a quarter of V4-Flash; persistent cache down to roughly an eighth; a causal encoder-decoder split that halves prefill work. The efficiency is architectural, not promotional.
The report also contains the most important product decision in the pack. DeepSeek trained an explicit reasoning-effort scalar, 1 to 100, into the system prompt, with a length penalty that decays exponentially as effort rises. The API exposes max, high, and low — effort 100, 75, 50. Effort from 25 to 100 lifts average Pass@1 across eight reasoning benchmarks from 67.1% to 76.3%, at about 2.5x the output tokens. DeepSeek's own advice: save max for the hardest tasks.
Which reframes the "most verbose model" finding. Artificial Analysis measured the max setting. That is the rightmost point on a cost-quality curve the model was explicitly designed to slide along. Verbosity became a dial.
Now the strategic consequence, which is where this stops being an accounting story. As open-weight competition compresses list prices toward marginal cost — Vercel's gateway data already shows cost per token at 43.3% of April levels while aggregate spend keeps rising — price stops being a differentiator anyone can hold. What remains scarce is tokens per completed task.
That makes the real frontier effort-aware length penalties, reasoning-effort controls, and cache compression: the machinery that decides how much a model says before it stops. Labs that win the next phase won't have the cheapest rate card. They'll have the model that stops talking.
And the frontier premium survives only on one condition: the expensive models must justify their verbosity. A 21% higher cost per coding task is defensible if the extra tokens buy capability a cheaper model can't reach. It is not defensible if the extra tokens are a spiral — 45 steps and 65 model calls for a $9.84 answer that a leaner model reached in 9 steps and 69 seconds for 71 cents.
The uncomfortable version for the premium labs: Bito tested 29 models and found 21 beaten outright by something cheaper scoring at least as well. deepseek-v4.1-flash scored 50.5 for $2.67 — 93% of the top score for 2% of the money. On simple, medium, and reasoning tiers, Bito's top eight cluster within about a point. Only the complex tier still separates them.
So the 80% cuts aren't a lie, exactly. They're a decoy. They move the number everyone negotiates over and leave untouched the number everyone pays.
Two threads stay live. Whether DeepSeek's token-efficiency gains translate into durable pricing pressure — the 50%+ consumption cut is the strongest evidence yet that they might, and it operates independent of any rate card. And whether the premium labs can convert verbosity from a liability into a priced feature, the way DeepSeek already has.
Watch the next release notes for effort dials and length penalties, not discounts. The discount is the part of the announcement that costs the lab nothing.