Week 39 - complete

API pricing and billing changes

September 21-27, 2026. Primary-source review completed September 28. Two priced model launches and two billing-policy changes met the weekly evidence gate.

4 dated changesOfficial sourcesAPI pricing

Price Change Summary

Tool	What Changed	Was	Now	Who This Affects
Claude Opus 5.5	New lower-priced Opus API model	Opus 5: $5 input, $25 output per million tokens	Opus 5.5: $4 input, $20 output per million tokens	Opus API teams evaluating migration
Claude API refusals	Billing resumed for three pre-output refusal categories	Not billed before output	Billed at the rate of the model that ran	Teams whose requests trigger bio, frontier_llm, or reasoning_extraction refusals
GPT-6 Sol and Luna	New priced API tiers	Not offered	Sol: $2 input, $10 output; Luna: $0.10 input, $0.50 output per million tokens	API teams selecting a new GPT-6 tier
xAI X Search	Billing changed from per-call to per-item rates	$5 per 1,000 calls	$5 per 1,000 posts; $10 per 1,000 profiles	Search-heavy agents

xAI

X Search: per-item billing replaces per-call pricing

Usage-sensitive

Starting September 21, 2026 at 12:00 PM PT (19:00 UTC), xAI changed X Search from $5 per 1,000 calls to $5 per 1,000 posts fetched and $10 per 1,000 user profiles fetched. Every returned post counts, including parent and quoted posts. Token charges are separate.

Wallet example

$0.50

100 posts returned

$1.00

100 user profiles returned

Both examples exclude token charges.

Who This Affects

Search-heavy agents should budget for returned posts and profiles as well as token usage.

Official source: xAI API pricing. A fresh raw capture retained September 28 explicitly states the September 21 effective time, the former $5-per-1,000-calls basis, and the new per-item rates.

Anthropic

Three pre-output refusal categories become billable

Billing resumed

On September 24, Anthropic resumed billing when a request refuses before producing output and stop_details.category is bio, frontier_llm, or reasoning_extraction. Those refusals are charged like any other request at the rate of the model that ran. Other pre-output refusal categories remain unbilled; mid-stream refusals were already billed.

Request state	Was	Now
Listed category, refusal before output	$0	Normal model rates
Other category, refusal before output	$0	$0

Wallet example

$0 to $4 for an Opus 5.5 request that consumes 1 million uncached input tokens, produces no output, and ends in one of the three newly billable refusal categories. Cache charges are excluded.

Who This Affects

Safety-sensitive API workloads should track these refusal categories because requests that previously ended before output at no charge can now incur the normal model bill.

Official sources: Anthropic release notes and Claude API pricing. Fresh raw captures were retained September 28.

Anthropic

Claude Opus 5.5 cuts the base Opus API bill by 20%

Lower rates

Anthropic launched Claude Opus 5.5 on September 22 at $4 input and $20 output per million tokens. Opus 5 remains $5 input and $25 output. Five-minute cache writes are $5, one-hour writes are $8, and cache reads are $0.20 per million tokens.

Model	Input	Output	Cache read
Opus 5.5	$4	$20	$0.20
Opus 5	$5	$25	$0.50

Wallet example

$24 for 1M uncached input plus 1M output on Opus 5.5, versus $30 on Opus 5. That is $6 less, or 20%.

Who This Affects

Teams running Opus 5 workloads can lower like-for-like base token costs by moving to Opus 5.5, but should test the documented API behavior changes before migration.

Official sources: Anthropic release notes, Opus 5.5 migration guide, and Claude API pricing. Independent raw captures were retained September 23.

OpenAI

GPT-6 Sol and Luna open a 20x output-price spread

New tiers

OpenAI added priced GPT-6 Sol and GPT-6 Luna API tiers on September 22. Standard short-context rates per million tokens are shown below.

Model	Input	Cached input	Cache write	Output
GPT-6 Sol	$2	$0.20	$2.50	$10
GPT-6 Luna	$0.10	$0.01	$0.125	$0.50

Wallet example

$12 vs $0.60 for 1M input plus 1M output. Luna is 95% cheaper for this token mix, but the models are not capability-equivalent.

Who This Affects

API teams choosing between the new GPT-6 tiers should model workload capability, token mix, and prompt length. Prompts above 272K input tokens use higher long-context rates.

Official sources: OpenAI API changelog, GPT-6 Sol model page, GPT-6 Luna model page, and OpenAI API pricing. Independent raw captures were retained September 23.

Switch watch

Do not compare price alone

Four practical checks before changing production traffic.

Refusal billing

Log stop categories

Measure newly billable pre-output refusals before forecasting the change.

Search volume

Count returned items

Measure posts and profiles returned per X Search call, plus token use.

Migration

Test Opus behavior

Review the four documented breaking changes before moving workloads to Opus 5.5.

Context

Separate long prompts

Model GPT-6 costs again when prompts cross the 272K-token threshold.

Method and limits

This is the latest complete briefing for September 21-27, 2026. We reviewed official Anthropic, OpenAI, xAI, Google Gemini API, GitHub, Cursor, and Notion pricing or release surfaces through September 27, with a final primary-source retrieval on September 28. The two model launches and both billing changes have dated Week 39 primary-source support. No additional pure pricing change met the evidence gate.

Read the prior complete briefing, Week 38 or browse all pricing briefings.