contracts, meters, provenance
self.md radar — 2026-04-28
The useful shift today is that the AI stack is being fenced in from three sides at once: contract terms, usage meters, and even the training corpus itself.
The opening signal is the rewritten Microsoft–OpenAI partnership, which sets new commercial fences. From there, the meter itself moves into view: GitHub, Anthropic, and DeepSeek all repriced agentic usage in ways users will feel. The closer is talkie, the strange but clean example of what happens when the corpus boundary becomes literal.
1. OpenAI and Microsoft rewrite the grand bargain
sources:
what happened: OpenAI and Microsoft published a restructured partnership. OpenAI can now serve products across any cloud provider, while Microsoft remains the primary cloud and gets first-ship rights on Azure unless it cannot or chooses not to support a given capability. Microsoft keeps a license to OpenAI IP through 2032 but it is now non-exclusive, no longer pays revenue share to OpenAI, and continues to receive revenue share from OpenAI through 2030 at the same percentage but capped. Simon’s audit notes that the old public AGI escape-hatch clause has effectively disappeared from the language around the deal.
why this matters: The contract layer just hardened: the parts of the stack that used to be vibes (AGI clause, exclusivity, mutual revenue share) are now bounded numbers and dates. For self.md the merge implication is that “OpenAI” and “Microsoft AI” are increasingly distinct vendors with their own surface area to track.
2. Coding agents reprice in public
sources:
- GitHub billing announcement
- GitHub pricing table
- Claude Code models and limits
- Claude extra usage
- DeepSeek pricing
what happened: GitHub announced that starting June 1, Copilot usage will consume GitHub AI Credits, and the annual-plan multipliers jumped sharply for Claude models: Opus 4.6 moves from 3x to 27x, Sonnet 4.5 from 1x to 6x. Anthropic’s Claude Code docs now frame Sonnet as the default and call out that Opus uses meaningfully more quota, while the extra-usage docs make the overflow path explicit: once included limits run out, paid users can continue at standard API pricing with spending caps. DeepSeek published a 75% discount on V4-Pro running through 2026-05-31 15:59 UTC.
why this matters: The meter is no longer background plumbing; multipliers, defaults, and overflow rules are now part of the user-facing surface of coding agents. For self.md this means tool choice and model choice need to be tracked against price tiers, not just capability.
3. talkie ships a 1930-cutoff base model
sources:
what happened: The talkie team released talkie-1930-13b-base, a 13B model trained on 260B tokens of pre-1931 English text. The 1930 cutoff is deliberate because that is the U.S. public-domain boundary, and the chat variant was post-trained without modern chat transcripts or off-the-shelf instruction-response data. The team reports the model can already handle small few-shot Python tasks despite having no modern web or code in its training mix.
why this matters: This is partly a science project about generalization and contamination, but it is also a clean illustration of provenance becoming a model-design choice rather than a downstream compliance footnote. For self.md the take is that “what was in the corpus” is starting to look like a first-class spec field, alongside parameter count and context length.
supporting links
- pip 26.1 — lockfiles and dependency cooldowns push the same boundary-setting instinct into package management.
- Memory Guardian — memory for agents is starting to come with explicit governance and decay rules.
- A2A-Agent-Action-Guard — safe tool use is getting split into its own control layer.
- CopilotKit — dedicated frontend stacks for agent sessions look more inevitable once those sessions are long, stateful, and billable.
left on the table
- Obsidian append app — same notes-layer wave already covered on Apr 23 and Apr 25; weaker delta.
- VibeVoice — strong repo, but same voice/local-model lane as Qwen TTS from Apr 23.
- GitHub Copilot 9x price increase thread — same pricing story, weaker than the official docs.
- Anthropic Opus paywall thread — same story as the support docs, just noisier.
- allowed repeat: the GitHub / Anthropic / DeepSeek pricing signal returns after Apr 22’s workload-tiered coding subscriptions, allowed because the delta this time is official policy and pricing movement across multiple vendors, not another round of complaints about expensive coding agents.