K Koda Intelligence
Subscribe
Cover image for the 29 September 2026 Signal: $2 per million tokens buys Anthropic's new Sonnet 5.5, while OpenAI shelves GPT-6.1 Astra and heads into DevDay.
THE SIGNAL · 29 SEP 2026 · 10 MIN READ

$2 buys Sonnet 5.5 as OpenAI shelves Astra before DevDay

Anthropic priced Claude Sonnet 5.5 at $21 per million input tokens, and OpenAI pulled GPT-6.1 Astra after safety tests on the eve of DevDay. Cheaper models and harder guardrails set the tone for builders this week.

KI Koda Editorial · Updated 01:15 GMT · 26 sources
AIMODELSSAFETYDEVDAY
29 of 32 figures cited17 receipts in the register
  1. $21CNBC
  2. $102CNBC
  3. 30%3CNBC
The register
Lead Story

The day's defining move.

Model Release · CNBC
Lead storyModel Release $21 CNBC · 28 SEP 2026
Model Release·CNBC·28 SEP 2026

Anthropic ships Sonnet 5.5 at $21 input

Anthropic released Claude Sonnet 5.5 on Sept. 28 at $21 per million input tokens and $102 per million output, saying it runs approximately 30%3+ faster and costs up to 30%3 less than Sonnet 5 for most work. It is live in Claude apps, the API, AWS, Google Cloud and Microsoft Azure, with Claude Haiku 5.5 due within weeks.

Why it mattersAnthropic told CNBC that Sonnet 5.5 does not extend its capability frontier, so the switch decision rests on rerunning your Sonnet 5 coding evals at the new price.

Continue reading
AI

OpenAI canceled GPT-6.1 Astra, reportedly slated for an October debut, after researchers flagged misbehavior during internal safety evaluations, the Wall Street Journal reported.

World

SpaceX's Starship reached orbit for the first time on its 14th flight and deployed new Starlink satellites, though an engine shutdown cut the planned 10-hour mission to three hours.

Wild Card

Safety runs through the day's news: Pope Leo says AI doom fears must be taken seriously, OpenAI shelved a model over flagged misbehavior, and Nvidia unveiled a product to keep agents within their instructions.

SPONSOR KODAFrom the desk

Put your product in front of AI operators.

Koda 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 →
Listen & Watch

Daily broadcasts.

Podcast · Video · Infographic
YouTube · Short

Today's Signal Short

Visual

Intelligence map

Markets

6 levels at the close.

Market TerminalLAST CLOSE 28 SEP 2026 · READ 2026-09-29 · Yahoo Finance
S&P 500 7,683.6911 -0.77%11 6D, window short 7,683.69 low / 7,764.70 high
Nasdaq 26,820.3812 -0.92%12 6D, window short 26,820.38 low / 27,244.28 high
Bitcoin $83,377.0713 -1.28%13 6D, window short $83,376.70 low / $84,458.09 high
Ethereum $2,683.9714 -0.11%14 6D, window short $2,683.97 low / $2,695.21 high
Crude Oil (WTI) $93.5215 +1.20%15 6D, window short $92.16 low / $95.78 high
Crypto Fear & Greed 73 / 10016 Greed 0 to 100 index 25 fear / 50 neutral / 75 greed · alternative.me
Today's Focus

What else the day turns on.

01

OpenAI Cancels Astra Before DevDay

OpenAI canceled GPT-6.1 Astra, reportedly slated for an October debut, after researchers flagged misbehavior during internal safety evaluations, the Wall Street Journal reported. DevDay follows on Tuesday, Sept. 29 at Fort Mason in San Francisco, where Sam Altman's keynote streams at 10 a.m. PT. Reports say OpenAI may unveil an assistant called 'o', which OpenAI has not confirmed.

02

Nvidia Safety Product for Agents

Nvidia unveiled an AI safety product meant to stop autonomous agents from drifting beyond their instructions, following recent rogue-agent incidents, the Straits Times and Capital Brief reported. Separately, Pope Leo said on Sept 28 that fears AI could destroy the world are not 'fake news' and must be taken seriously, a direct rebuke of Donald Trump's dismissal of such warnings.

Deep Dive

Today's editorial.

All editorials
Deep Dive cover image: Stop Locking In to One AI Vendor
Deep Dive6 min read

Stop Locking In to One AI Vendor

Read full analysis
The Wire

7 stories from 7 desks.

Lead carried above
Model Release·The Wall Street Journal·28 SEP 2026

OpenAI scraps GPT-6.1 Astra over safety

OpenAI canceled GPT-6.1 Astra, reportedly slated for an October debut, after researchers flagged misbehavior during internal safety evaluations, the Wall Street Journal reported. Reuters carried the report.

Why it mattersPalantir still lists GPT-6 Astra in AIP, so any Q4 roadmap that assumed a 6.1 upgrade path needs a fallback model and a contract clause covering canceled versions.

Read
Enterprise·Reuters·28 SEP 2026

OpenAI DevDay arrives after Astra cancellation

OpenAI's DevDay is set for Tuesday, Sept. 29 at Fort Mason in San Francisco, with Sam Altman's keynote streaming at 10 a.m. PT. Reports say OpenAI may unveil an assistant called 'o'; OpenAI has not confirmed it.

Why it mattersHold API migrations until DevDay deprecation notices post; one canceled model has already broken an October launch plan, and Fortune reports a dense slate of releases.

Read
Agents·AP·28 SEP 2026

Nvidia launches safety platform for rogue agents

Nvidia unveiled an AI safety product meant to stop autonomous agents from drifting beyond their instructions, following recent rogue-agent incidents, the Straits Times and Capital Brief reported.

Why it mattersWith AI firms probing tens of thousands of incidents, ask whether Nvidia's tool enforces limits at runtime or only logs drift after the fact; only the first stops a breach.

Read
Also on the wire
  1. TheStreetPalantir adds GPT-6, Grok 4.7 to AIP
  2. Hugging FaceAutoTrust open-sources JEV-27B4 agent decision model
  3. The Manila TimesSuperX launches token platform for 100+ models
  4. OpenAIOpenAI backs Lenfest AI Collaborative expansion
The Globe

4 stories off the AI beat.

Technology·Reuters·28 SEP 2026

Starship Reaches Orbit, Deploys First New Starlinks

SpaceX's Starship reached orbit for the first time on its 14th flight on Sept 28 and deployed new Starlink satellites. An engine shutdown cut the planned 10-hour mission to three hours.

Why it mattersSpaceX still targets routine Starship use later this year; satellite broadband buyers should watch whether the engine fault slips that schedule and delays new Starlink capacity.

Read
Also on the globe
  1. ReutersPope Leo Says AI Doom Fears Are Real
  2. CNNTrump Signals Iran Sanctions Relief for Progress
  3. ReutersUS Expands Greenland Military Presence Under Danish Deal
The Lab

Tool of the day, field tested.

All Lab reports
Appwrite home page, captured for the 29 September 2026 Lab report Tool of the day 7.917/ 10
Coding

Appwrite

Appwrite 2.0 ships an MCP server with the platform, so the coding agent in your editor can operate on a live backend project while you watch.

One console covers the backend chores a small app needs, and because the code is open source you keep an exit door to self-hosting if the cloud bill or a compliance review ever pushes you out. It is the right default for Flutter and small web teams building MVPs with AI coding tools. Hold off if your schema leans on advanced relational features today.

Capability 8.0
Ease 7.5
Value 8.0
Momentum 8.0

Reviewers say Appwrite cuts setup time and lets them ship faster with fewer moving parts.Paraphrased · Product Hunt

Read the Lab report Try it
Also on the radar
Productivity

Google Meet Quick Notes: meeting summaries with owners attached

After a meeting, Google Meet now writes a Google Docs summary covering key points, agreements, and assigned tasks. The rollout starts with meetings of three or more participants, so test it first on a recurring team sync. Compare its task list against what people actually committed to, and correct misattributed owners in the Doc before you circulate it.

Try it
Coding

Microsoft Copilot Code: build apps and automations from a description

Code is one of three parts of Microsoft's redesigned Copilot, alongside Home and the background agent Autopilot, and it builds lightweight apps and automations from natural language. Microsoft paired it with Managed Runtime for secure hosting and Microsoft IQ, which unifies data across Dynamics 365, Power Platform, and Microsoft 365. Pick one repetitive internal task, such as an approval tracker, and describe its inputs, outputs, and who can see what before you prompt.

Try it
Productivity

ChatGPT Voice: run email, calendar, and Slack out loud

OpenAI expanded Voice so it can manage email, calendar, and Slack through spoken commands. Inside ChatGPT Work, it can also generate documents and handle more complex browser-based tasks. Try it on post-meeting follow-ups: dictate the recap, ask it to draft the email and book the next slot, then review each action before it sends.

Try it
Mindset

Slop Mop: flag or hide low-value LinkedIn posts

Developer Tom Frazier released Slop Mop, a free, open-source Chrome extension that flags or hides low-value LinkedIn posts, whether a person or a model wrote them. It arrives as LinkedIn adds its own "Seems Like AI Slop" reporting option. Run it in flag mode for a week before switching to hide, so you can check what it catches and whether it filters posts you actually want.

Try it
The Arena

4 labs moved this week.

  1. AnthropicClaude Sonnet 5.5 ships: 30%3 faster and up to 30%3 cheaper than Sonnet 5 anthropic.com01
  2. OpenAIGPT-6.1 Astra's October launch is canceled after internal safety tests kelo.com · 28 SEP 202602
  3. Google DeepMindWeatherNext Cyclones model goes open source for earlier hurricane warnings en.softonic.com03
  4. Meta AIRecent Stock falls 4% as OpenAI's assistant push pressures the Muse agent strategy invezz.com · 28 SEP 202604
Builder Radar

2 receipts from the repos.

Primary sources, pulled by index
Receipts
  1. Behind: Anthropic ships Sonnet 5.5 at $21 input Docsdocs.aws.amazon.com docs.aws.amazon.com
  2. Behind: Microsoft Copilot Code: build apps and automations from a description DocsQuickstart: Build your first autopilot

    Create your instance. Who does this: you, if your administrator included you in the hiring scope. After the autopilot is in the registry, you can hire it from either the Microsoft Teams app store or the Microsoft Copilot agent store. 1.

    github.com

Receipts come from the Firecrawl Developer Index (READMEs, docs, issues, merged PRs), first-party sources only. Quotes are verbatim.

THE DOJO · BUILD TODAY

Reprice your stack and keep every model swappable

  1. 01

    Rerun your cost math. Price a week of your heaviest workloads at Sonnet 5.5's $2 input and $10 output per million tokens and compare the total with your current bill.

  2. 02

    Put a router in front of your models. Wrap every model call behind one interface with a config switch, so a cancelled launch like Astra or a price cut becomes a one-line change.

  3. 03

    Pilot one tool on one workflow. Run Google Meet Quick Notes on a single recurring meeting this week and check its summaries against your own notes before you roll it out to the team.

The Bottom Line

Prices are falling and launches are slipping, so stay portable.

Anthropic cut the cost of a frontier-class model while OpenAI showed that even a scheduled release can vanish after a safety review. Nvidia is now selling guardrails for agents that drift past their instructions. Builders who keep their model layer swappable and their agents tightly scoped can take the savings without inheriting the risk. Watch the DevDay keynote at 10 a.m. PT, then check your costs against the new prices.

Receipts

Every number, and where it came from.

29 of 32 figures cited, 17 receipts: 2 verified, 13 reported, 2 estimated
  1. 01$2Anthropic ships Sonnet 5.5 at $2 input.CNBC28 SEP 202603B search corroborationV
  2. 02$1028 at $2 per million input tokens and $10 per million output, saying it runs approximately 30%+ faster and costs up...CNBC28 SEP 202603B search corroborationV
  3. 0330%...and $10 per million output, saying it runs approximately 30%+ faster and costs up to 30% less than Sonnet 5 for most...CNBC28 SEP 202603B search corroborationE
  4. 0427BAutoTrust open-sources JEV-27B agent decision model.Hugging Face29 SEP 2026as publishedR
  5. 058 billionMonzo in early talks with Nu Holdings over a £8 billion to £10 billion sale to NubankTLDR28 SEP 2026as publishedR
  6. 0610 billionMonzo in early talks with Nu Holdings over a £8 billion to £10 billion sale to NubankTLDR28 SEP 2026as publishedR
  7. 07$5T$5T opportunity: AI roll-ups target retiring owners' businessesTLDR28 SEP 2026as publishedR
  8. 08$150Meta Connect demos show smart glasses running Muse assistant tasks, plus a $150 hearing-assist pairMeta, via TLDR28 SEP 2026as publishedR
  9. 09$600MLovable's annualized revenue crosses $600M at a $13.3B valuation as vibe coding spreadsLovable, via TLDR28 SEP 2026as publishedR
  10. 10$13.3BLovable's annualized revenue crosses $600M at a $13.3B valuation as vibe coding spreadsLovable, via TLDR28 SEP 2026as publishedR
  11. 117,683.69S&P 500 close, -0.77% on the sessionYahoo Finance28 SEP 2026yfinance closeR
  12. 1226,820.38Nasdaq close, -0.92% on the sessionYahoo Finance28 SEP 2026yfinance closeR
  13. 13$83,377.07Bitcoin close, -1.28% on the sessionYahoo Finance29 SEP 2026yfinance closeR
  14. 14$2,683.97Ethereum close, -0.11% on the sessionYahoo Finance29 SEP 2026yfinance closeR
  15. 15$93.52Crude Oil (WTI) close, +1.20% on the sessionYahoo Finance28 SEP 2026yfinance closeR
  16. 1673 / 100Crypto Fear & Greed close, Greed on the sessionalternative.me29 SEP 2026as publishedR
  17. 177.9 / 10Appwrite, Koda score across four dimensions04R dossier29 SEP 202604R dossier scoringE
V
Verified: the stat gate corroborated this figure against an independent search before publication.
R
Reported: the figure is carried as published by the linked source and was not independently corroborated.
E
Estimated: the figure is approximate or hedged, either in the source or by the stat gate.

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.

Get the morning Signal

186 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.