Verdict, October 18, 2026. Price the finished task, not the sticker. OP: Haiku 5.5 at or under 100,000 prompt tokens, and Sonnet 5.5, GPT-6.1 Sol standard, or Gemini 4 Argon intro pricing when the cache hits. Mid: Sol Ultrafast, Opus 5.5 fast mode, Argon Priority, Haiku over 100,000 tokens, and Argon after November 30, 2026. OP or Mid has not tested these models hands-on.
What actually changed
Secondary writeups dated an October 15 launch for Opus 5.5 and Sonnet 5.5. The model pages do not. Opus 5.5 released September 22, 2026 at $4 / $20 per million tokens. Sonnet 5.5 released September 28 at $2 / $10. Sources: Opus 5.5 (opens in a new tab), Sonnet 5.5 (opens in a new tab).
PRICE October 7: Sonnet 5.5 cache reads fell from $0.20 to $0.10 per million. Writes and the $2 / $10 sticker stayed, and Haiku 5.5 launched. Source: release notes (opens in a new tab). The same day, Codex CLI 0.161.0 made GPT-6.1 Sol the default in the bundled and Bedrock catalogs. Source: Codex 0.161.0 (opens in a new tab). Sol had shipped September 29 at $2 input, $0.10 cached, $2.50 cache write, and $10 output per million, up to 272K input tokens. Source: changelog (opens in a new tab).
PRICE October 8: Ultrafast for gpt-6.1-sol with service_tier: "ultrafast", for all API users, at 6x standard. OpenAI says this shortens the gap between output tokens. That is their description, not a test here. Sources: changelog (opens in a new tab), GPT-6.1 Sol (opens in a new tab).
PRICE October 13: Argon launched at $2 / $10, cached input 95% off, through November 30, 2026, then $4 / $20. Source: Google (opens in a new tab). The pricing page starts the higher card on December 1, cache included: $0.10, then $0.20. Source: Gemini pricing (opens in a new tab).
PRICE October 15: Opus 5.5 fast mode left research preview, with no beta header and default limits on Start, Build, and Scale. Source: release notes (opens in a new tab). It is $8 / $40 per million across the window, including prompts over 200k tokens, and it does not change the September 22 card. Against Opus 5 ($5 / $25, cache reads $0.50), that card is 20% less on input and output and 60% less on cache reads. Source: pricing (opens in a new tab).
What decides the task
Windows sit near a million tokens: 1M on the Claude 5.5 models, 1,050,000 on Sol, and 1,048,576 shared by input, thinking, and output on Argon. Sources: Haiku 5.5 (opens in a new tab), GPT-6.1 Sol (opens in a new tab), Argon (opens in a new tab).
PRICE Output is 5x input on the standard cards below. Thinking bills as output: Claude even when the text is omitted, OpenAI as reasoning tokens, Argon inside the output price. Sources: Thinking (opens in a new tab), Reasoning (opens in a new tab), Gemini pricing (opens in a new tab).
PRICE Cache reads are the loop discount: 0.05x on Opus 5.5 and Sonnet 5.5, 0.1x on Haiku, 5% on Sol. A Claude 5-minute write is 1.25x input and a 1-hour write is 2x. Sol's write is 1.25x. Sources: Claude pricing (opens in a new tab), GPT-6.1 Sol (opens in a new tab). Argon has no explicit cache. From 4,096 tokens, an implicit hit is 95% off and a miss is full price. Source: Argon (opens in a new tab).
NERF Two cliffs reprice the whole request. Over 100,000 prompt tokens, cache included, Haiku pays $0.50 / $2.50 instead of $0.10 / $0.50. Opus 5.5 and Sonnet 5.5 keep one rate through 1M. Source: pricing (opens in a new tab). Over 272K input tokens, Sol pays 2x input and cache and 1.5x output. Source: pricing (opens in a new tab).
NERF VENDOR CLAIM Anthropic says Haiku 5.5's tokenizer, in use since Claude 4.7, counts the same text as approximately 30% more tokens than Haiku 4.5. The gap depends on the content. Sonnet 4.6 and earlier keep the old one. Sources: Haiku 5.5 (opens in a new tab), pricing (opens in a new tab). That 30% is not a rate against Sol or Argon.
PRICE Speed multiplies the step. Sol Fast is 2x. OpenAI renamed Priority to Fast on July 30, 2026. Ultrafast is 6x, including on the long card ($24 / $1.20 cached / $30 write / $90). Cache multipliers stack on Opus fast. US-only Claude inference is 1.1x from Claude 4.6 on. OpenAI adds 10% for regional processing on models released on or after March 5, 2026. Batch is half of standard on Claude, Sol, and Argon Batch and Flex. Argon Priority is 1.8x standard: $3.60 / $18 and cache $0.18 through November 30, then $7.20 / $36 and cache $0.36. VENDOR CLAIM Google describes that tier as lower latency. Sources: Claude pricing (opens in a new tab), OpenAI pricing (opens in a new tab), Gemini pricing (opens in a new tab), Priority (opens in a new tab).
An illustrative step
Not a test, and not a live meter. Assumed counts: 20,000 fresh input tokens, 70,000 cache reads, 8,000 output tokens with thinking included, and no cache write. The 90,000-token prompt keeps Haiku on the low card and Sol on the short card. Argon's 70,000 are an assumed hit. This compares rate cards, not tokenizers.
cost = fresh/1,000,000 × input + cached/1,000,000 × cache read + output/1,000,000 × output
Sonnet: 0.02 × $2 + 0.07 × $0.10 + 0.008 × $10 = $0.04 + $0.007 + $0.08 = $0.127. Haiku's low card: 0.02 × $0.10 + 0.07 × $0.01 + 0.008 × $0.50 = $0.0067. Other rows use the same formula. Ultrafast is 6 × $0.127 = $0.762.
| Model | Input | Cache read | Output | This step |
|---|---|---|---|---|
| Opus 5.5 standard | $4 | $0.20 | $20 | $0.254 |
| Sonnet 5.5 | $2 | $0.10 | $10 | $0.127 |
| Haiku 5.5, up to 100,000 | $0.10 | $0.01 | $0.50 | $0.0067 |
| GPT-6.1 Sol standard | $2 | $0.10 | $10 | $0.127 |
| GPT-6.1 Sol Ultrafast | $12 | $0.60 | $60 | $0.762 |
| Argon standard through Nov 30, 2026 | $2 | $0.10 | $10 | $0.127 |
| Argon Priority through Nov 30, 2026 | $3.60 | $0.18 | $18 | $0.2286 |
| Argon standard from Dec 1, 2026 | $4 | $0.20 | $20 | $0.254 |
USD per million tokens. Prices checked October 18, 2026. A 5-minute Opus write on the 70,000 tokens is $0.35, against $0.014 for the hit. GPT-6 Sol, the ChatGPT model, is also $2 / $10, but the short-card cache read is $0.20, not 6.1's $0.10. Sources: Claude (opens in a new tab), OpenAI (opens in a new tab), Gemini (opens in a new tab).
Who it is for
Haiku 5.5 is OP for routing, extraction, and cheap subagent calls at or under 100,000 prompt tokens, and Mid over that line. The 30% note is Anthropic's, versus Haiku 4.5 only. See Claude.
Sonnet 5.5 is OP when you are not buying speed. The $2 / $10 card does not expire. Mid when the job needed Opus.
Opus 5.5 is Mid on this step, and OP on cost only against Opus 5 or Ultrafast. Fast mode is Mid unless the minute is the product.
GPT-6.1 Sol is OP on standard, including as the Codex CLI 0.161.0 default, and Mid on Ultrafast or past 272K input tokens. Codex is the harness.
Argon is OP through November 30, 2026 on standard if the cache hits, Mid on Priority, and Mid from December 1. No free tier. Contributor is invite-only at Batch rates, and that row says content is used to improve Google's products. Source: Gemini pricing (opens in a new tab). See Gemini API.
No letter moves. Argon, GPT-6, and the Claude 5.5 family stay in A on the October 15 Frontier Models list. S stays locked. AI Coding Agents ranks harnesses.
Loadout
- OP: Haiku 5.5 at or under 100,000 prompt tokens, or Sonnet 5.5, Sol standard, or Argon intro on a real cache hit.
- Mid: Ultrafast, Opus fast, Argon Priority, Haiku over 100,000 tokens, and Argon after November 30.
- October 15 did not launch Opus 5.5 or Sonnet 5.5.
The weekly patch is on beehiiv: subscribe free (opens in a new tab).