Week 43

Caches last longer. A Mistral discount ends.

October 19-25, 2026 UTC. OpenAI adds 12-hour prompt caching at its own write rate, Mistral Large 4 returns to list price after its launch offer, and Mistral and ElevenLabs publish new speech rate cards.

One launch offer endedOfficial sourcesReviewed October 26
What changed for your wallet
Price Summary
ToolWhat ChangedWasNowImpactWho It Affects
OpenAI API12-hour prompt caching added October 2030-minute cache writes only, 1.25x input (Sol $2.50 per M)Optional 12-hour writes at 2.5x input (Sol $5 per M); reads unchangedCheaper reuseResponses API users who reuse long prompts hours apart
Mistral Large 4Launch offer ended between our October 20 and October 26 checks$0.68 / $2.09 per M tokens (launch offer)$1.36 / $4.18 per M tokens (list)2x per tokenMistral API users on Large 4
Mistral Voxtral 3New audio family in Public Preview from October 20Voxtral TTS $16 per M chars; Realtime $0.006 per minTTS 3 $13 per M chars; Realtime 3 $0.008 per minNew rate cardMistral speech API users planning to upgrade
Mistral GLM 5.3 FlashZ.ai model available on Mistral from October 22GLM 5.3 $1.40 / $4.40 per M tokensGLM 5.3 Flash $0.15 / $0.50 per M tokensNew rate cardMistral API users who want a cheaper GLM option
ElevenLabsEleven Realtime voice agent models launched October 19No Realtime rate cardMini $0.20 / $6.00, Max $1.80 / $9.00 per M tokens, plus $0.08 per extra connected minuteNew rate cardDevelopers building voice agents on ElevenLabs

All amounts are USD. One launch discount ended, one new caching option and three new rate cards: no existing plan raised its list price this week. Each row links to its full story below.

Past briefings: Week 42 · Briefing archive

OpenAI 12-hour caching: pay more once, rewrite less

On October 20, OpenAI added 12-hour prompt caching to the Responses API. Until now, a cached prompt prefix had a 30-minute minimum lifetime. If your next request came later, you paid to write the same prefix again.

The two lifetimes are priced differently. A 30-minute cache write costs 1.25x the input price; a 12-hour write costs 2.5x. These are total write rates, not surcharges on top of input. Cache reads cost the same for both lifetimes. On GPT-6.1 Sol short context, that means input $2, a 30-minute write $2.50, a 12-hour write $5 and a cache read $0.10 per million tokens. GPT-6 Astra 12-hour writes cost $25 and GPT-6 Luna $0.25 per million short-context tokens.

Example: a 1M-token Sol prompt reused six times across a day, two hours apart. With 30-minute caching every request writes again: 6 x $2.50 = $15.00. With 12-hour caching you write once and read five times: $5 + 5 x $0.10 = $5.50, a $9.50 saving. If your requests arrive within 30 minutes of each other, keep the default: one write plus five reads costs $3.00 at 30 minutes and $5.50 at 12 hours.

Who This Affects

Responses API users with long, stable instructions or reference material that is reused across gaps longer than 30 minutes. Choose the lifetime per breakpoint. See ChatGPT API rates.

Official October 20 OpenAI API changelog, prompt caching guide and API pricing. Sources retained October 26, 2026.

Mistral Large 4: the launch discount is over

Mistral released Large 4 on October 6 with launch pricing of "50% off for 2 weeks". Our October 20 check still showed the sale price of $0.68 input, $0.07 cached input and $2.09 output per million tokens. On October 26 the pricing page shows only the list price: $1.36, $0.14 and $4.18. Mistral did not publish an exact end date, so the switch happened between our two checks.

Every Large 4 token now costs twice as much. A month with 10M input and 2M output tokens went from $6.80 + $4.18 = $10.98 to $13.60 + $8.36 = $21.96, before caching and batch discounts.

Who This Affects

Anyone who started on Large 4 during the launch weeks and budgeted at the sale price. Re-run your cost estimate at list price. See Mistral API rates.

Official Mistral changelog (October 6 entry) and Mistral API pricing. Sources retained October 20 and October 26, 2026.

New rate cards: Voxtral 3, GLM 5.3 Flash and Eleven Realtime

Mistral Voxtral 3 (October 20). Four new speech models, all in Public Preview. Voxtral TTS 3 costs $13 per million characters, against $16 for Voxtral TTS: 10M characters cost $130 instead of $160. Transcription goes the other way. Voxtral Mini Transcribe 3 costs $0.0035 a minute against $0.003 for Mini Transcribe 2, and Voxtral Mini Transcribe Realtime 3 costs $0.008 a minute against $0.006. For 10,000 minutes that is $35 instead of $30, and $80 instead of $60 for realtime. Voxtral Small Transcribe 3 costs $0.004 a minute.

Z.ai GLM 5.3 Flash on Mistral (October 22). A Flash version of GLM 5.3 at $0.15 input, $0.03 cached input and $0.50 output per million tokens. GLM 5.3 stays at $1.40 / $0.14 / $4.40. For 10M input and 2M output tokens, Flash costs $2.50 against $22.80 for GLM 5.3.

ElevenLabs Eleven Realtime (October 19). Two speech-to-speech models for voice agents, rolling out gradually across workspaces. Eleven Realtime v1 Mini costs $0.20 input and $6.00 output per million tokens; v1 Max costs $1.80 and $9.00. Connected audio beyond the included minutes costs $0.08 a minute on both.

Who This Affects

Mistral and ElevenLabs API users building speech features. Compare per-minute rates before moving a transcription workload to Voxtral 3. See Mistral API rates and ElevenLabs plans.

Official Mistral changelog, Mistral API pricing, ElevenLabs changelog and ElevenLabs API pricing. Sources retained October 26, 2026.

Switch Watch

Three actions tied to actual billing.

Opportunity

Cache day-long prompts for $5

A 1M-token Sol prompt reused every two hours costs $15.00 a day with 30-minute caching and $5.50 with one 12-hour write. Switch only the stable part of the prompt.

Warning

Rebudget Mistral Large 4

The launch price is gone. A 10M input, 2M output month now costs $21.96 instead of $10.98. Check last month's bill against list price.

Warning

Test Voxtral 3 before you move

TTS 3 saves $3 per million characters, but realtime transcription costs $0.008 a minute instead of $0.006: $80 instead of $60 for 10,000 minutes.

Also checked this week

HeyGen Business avatar slots. HeyGen's plan guide now says Business includes 5 Custom Digital Twin slots by default, or 10 in some regions, depending on region and account creation date. Our page said 10 by default. The wording had already changed by our October 20 check and we missed it last week; it is corrected now. The Business price is unchanged at $149 a month. See HeyGen plans.

Scalenut annual offer. On October 26 Scalenut's pricing page showed its annual offer as 65% off Early Access: $21, $31 and $70 a month billed annually for Starter, Plus and Professional, with 3X limits. On October 20 the same offer read 60% off at $24, $36 and $80. Scalenut published no dated notice and labels them Early Access offers; monthly prices stay $59, $89 and $199. See Scalenut plans.

Mistral rates now listed. The Mistral API table now shows rates our page had left blank: Voxtral Mini Transcribe Realtime at $0.006 a minute, Voxtral Small at $0.004 a minute of audio and Mistral Embed at $0.10 per million tokens. Our Mistral page lists them.

Coming up. Lovable says its current Chat mode pricing, including the per-message price and the daily chat allowance, applies through October 31, 2026, and calls it a temporary offering subject to change; no new price is published yet. Kimi closes its K2 series models to new users on October 31 and plans to retire them by the end of December; existing users are not affected. Mistral retires GLM 5.2 on October 31 and points users to GLM 5.3 at the same price. Gemini 4 Argon moves from $2 / $10 to $4 / $20 per million tokens on December 1.

Reporting window: October 19-25, 2026 UTC. All 35 tools were re-checked against official pages on October 26. Where a vendor shows only local-currency prices from our location, the last verified USD price and its date stay in place.