On one day, three providers republished the decks for the tier agent and coding workloads route to. OpenAI put GPT-6 Sol at $2.00 / $10.00 and GPT-6 Luna at $0.10 / $0.50 per million tokens; Anthropic put Claude Opus 5.5 at $4 / $20, below Claude Opus 5's $5 / $25; Google made Gemini 3.8 Flash TTS generally available.
per 1M tokens · standard short context
−20% vs Claude Opus 5 at $5 / $25
OpenAI, Anthropic and Google · Sep 22, 2026
Prices per million tokens in USD on the provider’s own API, observed September 28, 2026. The Gemini TTS pair is text input / audio output; it is a different purchase unit from a chat model, not a cheaper chat model.
| Model | Provider | Released | Price (in / out) | Context | Max output | Day-one surface |
|---|---|---|---|---|---|---|
| GPT-6 Sol | OpenAI | Sep 22, 2026 | $2.00 / $10.00 | 1,050,000 | 128,000 | OpenAI API — Responses and Chat Completions |
| GPT-6 Luna | OpenAI | Sep 22, 2026 | $0.10 / $0.50 | 1,050,000 | 128,000 | OpenAI API — Responses and Chat Completions |
| Claude Opus 5.5 | Anthropic | Sep 22, 2026 | $4 / $20 | 1M | 128K | Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry, Claude Platform on AWS |
| Gemini 3.8 Flash TTS | Sep 22, 2026 | $0.50 / $9.00 | — | — | Gemini API — Voices endpoint (text in / audio out) |
Sources: developers.openai.com · API changelog (GPT-6 Sol and Luna release and rates) · platform.claude.com · Claude Opus 5.5 (Opus 5.5 price, context, output) · ai.google.dev · Gemini API changelog (Gemini TTS availability)
Standard short context is the deck a rate card records. The other tiers are published, billable and usually invisible in a normalized registry row.
| Meter | Input | Cached input | Cache writes | Output | Note |
|---|---|---|---|---|---|
| Standard · short context (≤272K) | $2.00 | $0.20 | $2.50 | $10.00 | The deck a single per-deployment rate card records. |
| Standard · long context (>272K) | $4.00 | $0.40 | $5.00 | $15.00 | 2x input and cache, 1.5x output, applied to the whole request. |
| Batch / Flex | $1.00 | $0.10 | $1.25 | $5.00 | 50% of Standard on both legs. |
| Fast mode | $4.00 | $0.40 | $5.00 | $20.00 | 2x the applicable rate. Renamed from Priority Processing on Jul 30, 2026. |
Sources: developers.openai.com · GPT-6 Sol model page (context, output, long-context rule, batch / flex / fast multipliers) · developers.openai.com · API pricing (deck rates) · developers.openai.com · API changelog (Luna at one twentieth of Sol)
Where the model ID alone does not tell you what your integration will do.
GPT-6 Sol accepts text and image input and returns text, with reasoning effort from none to max. Built-in tools and function calling live on the Responses API; Chat Completions supports function calling only when reasoning_effort is set to none. A Chat Completions integration that wants tools at any reasoning level is a migration, not a model swap. Fine-tuning is not supported.
Regional processing adds a 10% premium where it is available, and EU data residency for GPT-6 Sol and Luna is offered only with Standard processing — not with Batch, Flex or Fast mode. A long-context EU route is therefore a standard-tier purchase by construction.
On September 25, 2026 OpenAI shipped a fix for an image-encoding bug that degraded image understanding in GPT-6 Sol and GPT-6 Luna, and it recommends rerunning evaluations and retrying affected workflows where image inputs are used. If you benchmarked these models between September 22 and 25 on visual tasks, the number is stale.
The registry has carried a model named gpt-5.6-sol for weeks. The new model is gpt-6-sol — a different ID, at a different price.
| Model ID | Standard input | Standard output | Status of the deck |
|---|---|---|---|
| gpt-5.6-sol | $4.00 | $20.00 | Promotional — “available at least through November 21, 2026” |
| gpt-6-sol | $2.00 | $10.00 | Standard deck on the September 22, 2026 release |
The new model’s standard input rate is half the promotional rate of the older one. Depending on which ID is actually called, a “Sol” line in a budget can be wrong by 2x in either direction — and the older ID’s deck is on a clock, so the two will not converge on their own.
Microsoft’s retail price list publishes the 5.6 meters under the product family Azure OpenAI GPT5 and the new ones under Azure OpenAI GPT6, with meter names 5.6-sol … and 6-sol …. The distinction is real in the meter data even where a deployment name is not.
Two Microsoft surfaces currently disagree about GPT-6 Sol on Azure. The meters are the machine-readable source; the pricing page is the human one.
The Azure OpenAI pricing page carries the notice that “GPT-6 Sol and Luna prices are currently in processing for publishing on this page”, and every GPT-6 Sol and Luna price cell on it is empty. The Azure Retail Prices API already returns the Azure OpenAI GPT6 meters for 6-sol in Standard, Data Zone and Priority Processing, in both short and long context.
| Azure meter (Global) | Azure rate | OpenAI direct | Verdict |
|---|---|---|---|
| 6-sol ShortCo Inp Std Gl | $2.00 | $2.00 | Standard short-context input — matches direct |
| 6-sol ShortCo Cd Inp Std Gl | $0.20 | $0.20 | Cached input — matches direct |
| 6-sol ShortCo Cd Wr Std Gl | $2.50 | $2.50 | Cache writes — matches direct |
| 6-sol ShortCo Opt Std Gl | $10.00 | $10.00 | Standard short-context output — matches direct |
| 6-sol LongCo Inp Std Gl | $4.00 | $4.00 | Long-context input — matches direct |
| 6-sol LongCo Opt Std Gl | $15.00 | $15.00 | Long-context output — matches direct |
| 6-sol ShortCo Inp PP Gl | $4.00 | $4.00 (Fast mode) | Azure Priority Processing = Fast mode |
| 6-sol ShortCo Opt PP Gl | $20.00 | $20.00 (Fast mode) | Azure Priority Processing = Fast mode |
Sources: prices.azure.com · Azure Retail Prices API · developers.openai.com · API pricing. Microsoft Foundry’s model catalog already lists gpt-6-sol and gpt-6-luna with a 2026-09-22 version date, and notes that the GPT-6 family requires a quota request on some subscription tiers: learn.microsoft.com · Foundry model catalog
Anthropic moved the Opus line down a price tier on the same day, and attached request-shape changes that fail loudly.
claude-opus-5-5 is priced at $4 / $20 per MTok against Claude Opus 5 at $5 / $25, with a 1M-token context window, 128K maximum output and adaptive thinking always on. Cache writes are $5 (5-minute) and $8 (1-hour) per MTok, and a cache read costs $0.20 — 5% of input rather than the 10% charged on most Claude models. For long agent loops with a stable system prompt, that cache line moves the effective rate more than the headline does.
Fast mode (research preview) prices Claude Opus 5.5 at $8 / $40 per MTok, against $10 / $50 for Claude Opus 5 and Opus 4.8. It applies across the full context window including requests above 200k input tokens, and it is available on the Anthropic first-party API only — not on Claude Platform on AWS or partner-operated platforms, and not with the Batch API.
Claude Opus 5.5 shipped simultaneously on the Claude API, Amazon Bedrock, Claude Platform on AWS, Google Cloud and Microsoft Foundry, so the release is reachable without leaving an existing cloud contract. Its listed retirement is not sooner than September 22, 2027.
A new purchase unit, and a capacity control that is not a deprecation.
Google made both TTS models generally available together with a Gemini API Voices endpoint, plus voice design, voice replication with consent verification and a library of 150+ voices. The flagship TTS model is priced at $0.50 per million text input tokens and $9.00 per million audio output tokens through December 31, 2026 — about $0.00225 per ten seconds of audio — and both rates double on January 1, 2027 ($1.00 / $18.00).
On September 18, 2026 Google limited access to the Gemini 2.5 models to users who had already used them, while stating that those models are not deprecated and will continue to be served, pointing new projects at 3.5 Flash-Lite or 3.8 Flash. A model can therefore be present, priced and undeprecated while no longer being provisionable for a new project — a distinction that a lifecycle-only view misses.
This report covers a new model and a new deck. The earlier one covers one model’s meter spread, its promotional clock and the registry records that disagree with it.
Read GPT-5.6 Sol: one model, six meters, and a price that expires for the $5.00 / $30.00 pre-promotional deck, the registry’s alias records and the residency uplift on that model. This report covers the September 22, 2026 releases: two new GPT-6 models, an Opus price tier change and Google’s TTS pair.
All official provider pages observed September 28, 2026 (Europe/Paris). Azure meter rates read from the Azure Retail Prices API for the Azure OpenAI GPT6 product family on the same date.
Related live views: the OpenAI vs Azure AI deployment comparison and the GPT-5.6 Sol price-deck report render the current catalog records.
Deployment intelligence, explained
Two things move independently when a provider ships a model: the model ID and the price deck attached to it. On September 22, 2026 both OpenAI and Anthropic republished the deck for the tier where coding and agent traffic lands — and OpenAI reused a name, “Sol”, that already belonged to a different model at a different price.
A published deck is a set of purchases, not a single rate. Standard, Batch, Flex, Fast mode, short and long context, and data residency are separate meters on the same model, so the headline rate and the rate a workload actually bills on can differ several-fold before residency is counted. Read the meter your traffic lands on, then the promotional clock attached to it.
We use technologies necessary for the website to operate. With your consent, we use PostHog and Google Analytics to measure pseudonymised audience data. Refusal does not affect access to the website.