Anthropic cancels Sonnet 5's price increase — and the week's biggest release has no spec sheet

1 month ago

The state of the board, for the week ending Friday, August 21st. What moved. Three releases, and the biggest one shipped without a spec sheet. What it costs. A scheduled price increase just got cancelled. And what to use, strictly by the numbers. This is Model Watch.

Ask

Ask about this presentation

Answers are generated from this presentation.

Chapters

  1. 0:00Model Watch — State of the Board
  2. 0:14The week on the board
  3. 0:40What it costs
  4. 1:06The board — price against score
  5. 1:36What to use
  6. 1:56Sources
Show transcript

Model Watch — State of the Board

Model Watch
State of the Board · week ending Friday · August 21
01What moved — three releases, one without a spec sheet
02What it costs — a scheduled price increase, cancelled
03What to use — strictly by the numbers
Model Watch · bisque.todayEvery release, measured

The state of the board, for the week ending Friday, August 21st. What moved. Three releases, and the biggest one shipped without a spec sheet. What it costs. A scheduled price increase just got cancelled. And what to use, strictly by the numbers. This is Model Watch.

The week on the board

What moved · the week on the board
Model Watch · bisque.todayFriday · August 21

The week on the board. Z A I shipped G L M five point two Turbo on Monday, its second release in a week, right behind G L M five point three. Google's Gemini three point seven Flash landed the Thursday before. But the headliner is not a model at all. It is a price. Anthropic's pricing page now says the scheduled September increase for Claude Sonnet five will not occur. Two dollars in, ten dollars out, per million tokens. That was introductory pricing. It is now just the price.

What it costs

What it costs · the pricing moves
The floor keeps dropping
$2 in · $10 out
Claude Sonnet 5 — increase cancelled
The scheduled September rise to $3 / $15 will not occur; the introductory price is now standard.
source: Anthropic pricing docs · Aug 21
$0.20 in · $1.20 out
GPT-5.6 Luna — down 80% since July 30
OpenAI's small-tier price after July's cut across the GPT-5.6 line.
as listed by CloudZero · Aug 2026
$0.14 in · $0.28 out
DeepSeek V4-Flash — the cheapest listed API
The current floor of the board, per million tokens.
as listed by CloudZero · Aug 2026
Model Watch · bisque.todayFriday · August 21

What it costs. Claude Sonnet five holds at two dollars in and ten dollars out. The scheduled rise to three and fifteen is cancelled. That is straight from Anthropic's pricing documentation. G P T five point six Luna lists at twenty cents in and a dollar twenty out, down eighty percent since OpenAI's July cut, as listed by CloudZero. And the cheapest listed A P I on the board is DeepSeek V four Flash, at fourteen cents in and twenty eight cents out. The floor keeps dropping.

The board — price against score

The board · Artificial Analysis Intelligence Index v4.1.1 (independent) × provider list price
59 61 63 $25 $30 $50 score · AA index ↑ output price · $ per Mtok →
Claude Opus 5 · 63 · $25 out
Claude Fable 5 · 62 · $50 out
GPT-5.6 Sol (max) · 61 · $30 out
GLM-5.2 Turbo · released Aug 17 · no published price, no published benchmarks — no numbers, no dot
Model Watch · bisque.todayFriday · August 21

The board itself. Three dots to start, all scored by Artificial Analysis, an independent index. Claude Opus five. Sixty three, at twenty five dollars per million output tokens. Claude Fable five. Sixty two, at fifty. G P T five point six Sol at max effort. Sixty one, at thirty. Read it plainly. Of the three, the highest score carries the lowest output price. And the open teal circle is G L M five point two Turbo. Z A I has published no prices and no benchmarks for it yet. On this board, no numbers means no dot.

What to use

What to use · by the numbers
Match the model to the job
Deep reasoning, hard agent workClaude Opus 5 — $5 in / $25 out, and the top score on the board at 63.
The default workhorseClaude Sonnet 5 — $2 in / $10 out, and that price is now permanent.
Volume and simple tasksHaiku 4.5 at $1 / $5, or GPT-5.6 Luna at $0.20 / $1.20 as listed.
Model Watch · bisque.todayFriday · August 21

So what do you run. For deep reasoning and hard agent work, Claude Opus five. Five dollars in, twenty five out, and the top score on the board. For the default workhorse, Claude Sonnet five. Two in, ten out, and that price is now permanent. For volume, Haiku four point five at one and five, or G P T five point six Luna at a fraction of that. Match the model to the job, not to the headline.

Sources

Sources
Every number traces to these
Claude pricingAnthropic pricing documentation, platform.claude.com, retrieved Aug 21 — Sonnet 5 increase cancelled; Opus 5, Fable 5, Haiku 4.5 rates.
Index scoresArtificial Analysis Intelligence Index v4.1.1, retrieved Aug 21 — independent.
Release datesLLM Gateway release timeline — GLM-5.2 Turbo Aug 17, GLM-5.3 Aug 14, Gemini 3.7 Flash Aug 13.
GPT-5.6 + DeepSeek pricesCloudZero LLM API pricing listings, Aug 2026 — provider-listed, not independently metered.
GLM-5.2 Turbo appears without specifications: Z.AI had published no prices or benchmarks for it as of Friday. It gets its dot when the numbers exist.
Model Watch · bisque.todayThe board updates Monday

Every number you just heard is on screen with its source, and the provider listed ones are labeled that way. That is the first board. It updates Monday morning.