Shifts · updated 9 Oct 2026
The ground keeps moving. 32 shifts in 68 days.
New models, new limits, new owners, new terms: what changed for people who code with agents, since August. Each one is a reason not to tie your work to a single agent.
- 3 Aug Qwen3.8-Max
- 14 Aug Qwen3.8-27B, small enough to run yourself
- 14 Aug SpaceX completes its $60 billion purchase of Cursor
- 14 Aug GLM-5.3, an open coding model
- 29 Aug OpenAI ends Cursor’s direct model access
- 31 Aug DeepSeek V4 Flash Vision weights, under MIT
- 1 Sept Copilot Business and Enterprise lose their extra credits
- 3 Sept GPT-6 Astra
- 10 Sept New ChatGPT Pro sign-ups are paused
- 14 Sept Claude Code’s weekly limits drop about 17%
- 16 Sept GPT-5.5 leaves ChatGPT and Codex on 14 October
- 18 Sept Claude Code reads AGENTS.md
- 22 Sept GPT-6 Sol and Luna, the same afternoon
- 22 Sept Claude Opus 5.5
- 25 Sept Codex outage, then early limit errors
- 29 Sept GPT-6.1 Sol, a week after GPT-6 Sol
- 29 Sept ChatGPT Pro reopens with less usage
- 30 Sept Gemini 4 Argon is announced
- 30 Sept Claude Sonnet 4.5 is deprecated
- 1 Oct Copilot Business and Enterprise seats are paid up front
- 1 Oct gpt-5.3-codex and gpt-5.1 leave the API in April 2027
- 1 Oct gpt-5.4-cyber removed from the API
- 2 Oct Four models deprecated across Copilot
- 5 Oct Reflection announces Beam, an open-weight model
- 6 Oct Mistral Large 4, a trillion-parameter model
- 7 Oct Local sandboxing for Copilot is generally available
- 7 Oct Some organisations wrongly paused at spend limit
- 7 Oct Claude Haiku 5.5 in GitHub Copilot
- 7 Oct Claude Haiku 5.5 in the API
- 8 Oct Gemini, a universal agent for the workplace
- 9 Oct Revised usage policy takes effect 12 November
- 9 Oct Dynamic workflows for Managed Agents, in beta
October 202613
Anthropic
Claude Platform release notesTools
Dynamic workflows for Managed Agents, in beta
An agent can write a program that runs many agents in phases and combines their results, for work like reviewing hundreds of documents.
Anthropic
The RegisterAccess
Revised usage policy takes effect 12 November
It bans sustained, needless abuse of Claude and tightens the rules on influence operations. Anthropic says the abuse rule targets only extreme cases.
Google
CBS NewsTools
Gemini, a universal agent for the workplace
Works in Gmail, Docs, Drive and Sheets, generates and runs code, and can start sub-agents that work for days. Availability is not yet stated.
Anthropic
Claude Platform release notesNew model
Claude Haiku 5.5 in the API
A 1M-token context window and 128k output. Code written for Haiku 4.5 can return 400 errors on it, so check before switching.
Anthropic
GitHub ChangelogNew model
Claude Haiku 5.5 in GitHub Copilot
Anthropic’s lightweight model is generally available in Copilot on Pro and up, billed at list price under usage-based billing.
Anthropic
Claude StatusOutage
Some organisations wrongly paused at spend limit
Requests were refused across the API, Claude.ai, Claude Code and Cowork. The incident is marked resolved.
GitHub
GitHub ChangelogTools
Local sandboxing for Copilot is generally available
Commands that Copilot agents start get limited access to files, network and credentials, under policies admins can lock. No extra cost.
Mistral
TechCrunchNew model
Mistral Large 4, a trillion-parameter model
A preview of Mistral’s largest model, reachable through a guardrail endpoint for now. Weights follow about three weeks later, after safety testing.
Reflection AI
AI BusinessOpen model
Reflection announces Beam, an open-weight model
501 billion parameters, 23 billion active, with a 1M-token window, for coding and agent work. Weights are due later this month; the claims are unverified.
GitHub
GitHub ChangelogAccess
Four models deprecated across Copilot
Gemini 3.5 Flash, Gemini 3.6 Flash, Kimi K2.7 Code and Claude Opus 4.7. Replacements named are Gemini 3.8 Flash, Kimi K3 and Claude Opus 5.5.
OpenAI
OpenAI API deprecationsAccess
gpt-5.4-cyber removed from the API
Removed on the date OpenAI gave in September. Its notice points users to the most capable cyber model available to them.
OpenAI
OpenAI API deprecationsAccess
gpt-5.3-codex and gpt-5.1 leave the API in April 2027
Also gpt-5.4-nano. Removal on 1 April 2027, with six months’ notice. The replacements are GPT-6 models.
GitHub
GitHub ChangelogPricing
Copilot Business and Enterprise seats are paid up front
For card and PayPal customers, each assigned seat is charged at the start of the cycle and needs payment before access. Invoice billing is not covered.
September 202613
Anthropic
Claude Platform docsAccess
Claude Sonnet 4.5 is deprecated
It retires from the Claude API on 30 November 2026. Anthropic’s replacement is Sonnet 5.5.
Google
Yahoo FinanceNew model
Gemini 4 Argon is announced
Google’s first Gemini 4 frontier model, built for long-horizon coding. Cyber defenders get it first; paid API access comes later.
OpenAI
Kingy AIPricing
ChatGPT Pro reopens with less usage
The $200 plan comes back with roughly half its old allowance, beside a new $500 tier. Existing subscribers keep the old allowance until 29 October.
OpenAI
Unite.AINew model
GPT-6.1 Sol, a week after GPT-6 Sol
Near GPT-6 Astra on agentic coding at a fifth of its price, in Codex and the API from DevDay.
OpenAI
OpenAI Developer CommunityOutage
Codex outage, then early limit errors
Errors on ChatGPT and Codex ran into the next day, and Pro users hit usage walls without notice. Paid limits were reset early afterwards.
Anthropic
BetaNewsNew model
Claude Opus 5.5
A cheaper model for long agent work. Anthropic says it matches Fable 5.1 on most tasks and costs 40% less to run than Opus 5.
OpenAI
Crowdfund InsiderNew model
GPT-6 Sol and Luna, the same afternoon
Faster, cheaper models under GPT-6 Astra, in Codex and the API, released hours after Claude Opus 5.5.
Anthropic
Enterprise DNATools
Claude Code reads AGENTS.md
Claude Code now reads AGENTS.md in projects that have no CLAUDE.md.
OpenAI
BusinessTodayAccess
GPT-5.5 leaves ChatGPT and Codex on 14 October
Retired from ChatGPT, ChatGPT Work and Codex on every plan. The API is not affected, but Codex settings that pin it need to move first.
Anthropic
Android HeadlinesLimits
Claude Code’s weekly limits drop about 17%
A 50% summer boost ended, replaced by a permanent 25% raise over the old baseline. Announced on 30 August.
OpenAI
TechCrunchAccess
New ChatGPT Pro sign-ups are paused
Demand for GPT-6 Astra outran capacity. People already on Pro kept it; nobody new could join.
OpenAI
WikipediaNew model
GPT-6 Astra
OpenAI’s new flagship for long coding and computer-use work: a preview on 3 September, generally available the next day.
GitHub
The GitHub BlogPricing
Copilot Business and Enterprise lose their extra credits
The promotional credits that came with June’s move to usage-based billing ended; included usage went back to standard amounts.
August 20266
DeepSeek
Digital AppliedOpen model
DeepSeek V4 Flash Vision weights, under MIT
A multimodal model in the V4 family, free to download and run.
OpenAI
CNBCAccess
OpenAI ends Cursor’s direct model access
After SpaceX bought Cursor, OpenAI said Cursor’s direct access to its models ends on 12 November, and its upcoming models won’t go there.
Z.ai
MarkTechPostOpen model
GLM-5.3, an open coding model
Z.ai’s coding model for long engineering work, with open weights following two weeks later.
Cursor
TechzineAccess
SpaceX completes its $60 billion purchase of Cursor
One of the most used AI editors now belongs to a company with its own models.
Alibaba
EmergentOpen model
Qwen3.8-27B, small enough to run yourself
Open weights under Apache 2.0, eleven days after the much larger Qwen3.8-Max.
Alibaba
DataNorthOpen model
Qwen3.8-Max
Alibaba’s new flagship: a 2.4-trillion-parameter model with a million-token window.
We link the reporting; we don’t host it. Missing a shift, or one we got wrong? Tell us on GitHub.