Anthropic and OpenAI both cut the cost of frontier-class output on Sept. 22, with Claude Opus 5.5 claiming 40%1 lower running costs and GPT-6 Sol and Luna claiming 50%2 lower API prices. Meanwhile the chips underneath got more expensive.
OpenAI resets API pricing with Sol and Luna OpenAI launched GPT-6 Sol and Luna on Sept.
Xiaomi open-sources the MiMo-V2.6 family Xiaomi announced MiMo-V2.6 on Sept.
AMD crosses a trillion on AI-chip demand Quartz reported AMD first reached a $1 trillion 4 market capitalization on Sept.
OpenAI launched GPT-6 Sol and Luna on Sept. 22 for Codex and ChatGPT Work. The claimed 50%2 API price reduction is measured against the GPT-5.6 promotional rates they replace, so read the baseline before you re-forecast spend.
Source: techcrunch.comXiaomi announced MiMo-V2.6 on Sept. 22, including Pro and Flash. It says Pro-UltraSpeed generates output up to 20x3 faster than Pro at the same quality, a claim worth testing on your own prompts before you trust it.
Source: venturebeat.comQuartz reported AMD first reached a $1 trillion4 market capitalization on Sept. 21, with shares up as much as 10%5 in a five-session rally. Inference gets cheaper per token while the silicon behind it gets repriced upward.
Source: reuters.comKoda publishes a fact-checked AI intelligence briefing every morning: site, email, podcast, and video. Founding sponsors lock today's rates for 6 months as the audience compounds.
View the media kit →Anthropic launched Claude Opus 5.5 on Sept. 22, saying it matches Fable 5.1 on most work and costs 40%1 less to run than Opus 5.
Why it mattersEnterprise teams should test the claimed 40%1 running-cost reduction on their own agent workloads before revising inference budgets.
Continue readingDonald Trump said US and Iranian delegations held three hours of talks at UN headquarters in New York, calling them productive and saying further meetings were planned.
AMD first reached a $1 trillion4 market capitalization on Sept. 21 as shares rose as much as 10%5, extending a five-session AI-chip rally, Quartz reported.
The performance pitches differ: Tencent says Hy Image 3.5 Preview matches Seedream 5.0 Pro, while Xiaomi claims Pro-UltraSpeed delivers up to 20x3 faster output than Pro at unchanged quality.
OpenAI launched GPT-6 Sol and Luna on Sept. 22 for Codex and ChatGPT Work. Its claimed 50%2 API-price reduction is measured against the GPT-5.6 promotional rates they replace, a different yardstick from Anthropic's running-cost comparison.
Source: techcrunch.comXiaomi announced the open-source MiMo-V2.6 family, including Pro and Flash, on Sept. 22. Its claim that Pro-UltraSpeed generates output up to 20x3 faster than Pro at the same quality makes speed, rather than improved quality, the stated distinction.
Source: venturebeat.com
Three AI Giants Cut Prices in One Day
Read full analysis
Tool of the day
5.716/ 10
Put the AI Doctor label aside: Juno's first test should be one bedtime reminder.
Habit tracking is the clearest reason to try Juno; its symptom advice is not a substitute for medical judgment. People building a basic wellness routine should give its reminders a seven-day trial with one low-risk habit. Adopt it only if those nudges help, and leave diagnosis and treatment to healthcare professionals.
The app feels simple to use and makes everyday health information easy to understand.Paraphrased · Google Play reviews, as summarized in the supplied research
Spoke captures important moments from video conversations, creates subtitled summaries and lets you edit the resulting video by editing text. Test it on one recording you have permission to use: select three key decisions, trim the transcript and check the subtitles against the original audio before sharing.
Try itThe New Black focuses on AI fashion design, although the supplied directory scrape does not document version-specific capabilities or pricing. Start with one garment brief specifying silhouette, fabric and color palette, then compare three concepts for consistency and use them as visual references rather than manufacturing specifications.
Try itWarp combines a terminal with AI assistance, but the supplied scrape contains an unrelated featured promotion rather than evidence of version-specific changes. Evaluate it in a disposable repository: ask for an explanation of one failing command, review the proposed fix and inspect every command before executing it.
Try itThis FutureTools entry points to Plexigen AI, described as generating videos with sound from text prompts or images; Future Tools is the directory label, not the generator's name. Test one image you own with a short motion brief, then check sound alignment, export quality and commercial-use terms before using the result publicly.
Try it[Windows][Computer Use] Works with GPT-6 Astra but unavailable with GPT-5.6 Sol, Terra, and Luna44174.Description. Amjadov What version of the Codex App are you using (From “About Codex” dialog)?.
github.comReceipts come from the Firecrawl Developer Index (READMEs, docs, issues, merged PRs), first-party sources only. The paper comes from the Firecrawl Research Index, arXiv, last 7 days, picked for today's lead. Quotes are verbatim.
Rebuild your cost model on real baselines. Anthropic's 40% figure is against Opus 5 and OpenAI's 50% is against GPT-5.6 promotional rates. Rerun your top three production prompts on Claude Opus 5.5 and GPT-6 Sol and price the actual token mix, not the headline.
Test the speed claim locally. Pull the open-source MiMo-V2.6 Pro and Flash builds and compare Pro-UltraSpeed output against Pro on your own eval set. A 20x throughput claim only matters if quality holds on your workload.
Wire one new tool into a live task today. Try Warp v2.0 for terminal-side agent work, Spoke for capture, and The New Black v1 for image generation, then benchmark Tencent's Hy Image 3.5 Preview against whatever you use now.
Three labs shipped cost or speed claims in a single day, and every one of them is measured against a baseline the vendor chose. That makes your own evals the only pricing authority that counts this week. AMD passing a trillion dollars on AI-chip demand is the other half of the trade: per-token costs fall while the hardware layer capturing that demand keeps compounding. Build for the cheaper inference, but assume the floor under it is still being repriced.
Method names the computation path behind the row, from a closed vocabulary: as published, yfinance close, 03B search corroboration, 04R dossier scoring, not recorded.
A figure is money, a percentage, a spelled magnitude, a multiple, a rate carrying its unit, a score over its denominator, a thousands separated number or an index close. Bare integers and years are not counted, so a phrase like "over 100 integrations" carries no mark. A chart's own scale line is the axis of the level above it and is counted once, with that level.
180 editions so far, one a day. Unsubscribe anytime.
Forward this to one operator you work with. Your referral link is in every email; milestones at koda.community/refer.