The week AI’s platform alliances got renegotiated in public

TL;DR

• OpenAI ended Microsoft cloud exclusivity, killed the AGI clause, and shipped GPT-5.5 on AWS Bedrock at API price parity within days.
• The Musk v. Altman trial is jury-advisory only; remedies sit with a judge who has already mocked the $134B damages figure.
• Anthropic launched Claude MCP connectors for Adobe, Autodesk, Blender and five others; a €240K Blender patronage triggered immediate artist backlash.
• Romain Huet confirmed no standalone GPT-5.5-Codex — the coding specialist track is folding back into the flagship line.
• Regulators are everywhere: EU pushing Gemini off Android, China blocking Meta’s Manus buy, OpenAI clearing FedRAMP Moderate.

Today’s news is about who gets to sit next to whom. OpenAI formally unbolted itself from Microsoft, shipped GPT-5.5 on AWS Bedrock the same week, and watched its old AGI clause quietly expire — a contract rewrite that ripples straight into the Musk v. Altman courtroom, where the “captured subsidiary” framing was the prosecution’s theory of the case until Monday morning. Anthropic, meanwhile, tried to buy its way into the creative-software stack with eight MCP connectors and a Blender Foundation patronage check, and got a revolt instead of a coronation.

Underneath the features, the round-ups tell the same story from the regulator’s chair: Brussels is trying to pry Gemini off Android, Beijing killed Meta’s Manus deal, and OpenAI just cleared FedRAMP Moderate to compete for federal workloads on Microsoft’s home turf. The platform alliances that defined the last three years of AI are being renegotiated in public — by contract, by court, and by competition authority.

OpenAI cuts the Microsoft cord — and the AGI clause dies with it

Source: openai-blog · published 2026-04-27

TL;DR

• OpenAI and Microsoft tore up exclusivity: Azure stays “first,” but OpenAI can now ship anywhere Microsoft can’t or won’t.
• GPT-5.5 and Codex landed on AWS Bedrock the same week, at price parity with OpenAI’s direct API.
• The AGI clause is gone — revenue is now “independent of OpenAI’s technology progress” through 2030/2032.
• TD Cowen sees Microsoft saving ~$700M in FY26 on killed revenue-share-out, but ~45% of MSFT’s AI backlog still rides on OpenAI.

One restructure, three forced moves

The OpenAI–Microsoft amendment isn’t a strategic pivot — it’s the cleanup after three external shoves arriving in the same quarter. The FTC’s April 8 AI Interoperability mandate made cloud exclusivity legally precarious by requiring hyperscalers to let enterprise customers port fine-tuned models across clouds without penalty ¹. Amazon’s reported $50B compute deal with OpenAI created a contractual collision Microsoft couldn’t paper over. And OpenAI needs IPO-ready optionality ahead of a rumored late-2026 listing at a >$1T valuation. The simultaneous Bedrock launch — GPT-5.5, GPT-5.4, Codex, and Managed Agents going live on AWS — is the proof the new terms are operational, not aspirational.

flowchart LR
    O[OpenAI models + Codex] -->|first launch| A[Azure]
    O -->|now permitted| B[AWS Bedrock]
    O -->|"any cloud if MSFT can't serve"| X[Other clouds]
    A -. "rev-share to MSFT, capped, through 2030" .-> M[Microsoft]
    M -. "IP license through 2032, non-exclusive" .-> O

The most-overlooked clause is the quietest: revenue obligations are now “independent of OpenAI’s technology progress,” which Simon Willison reads as the formal death of the AGI tripwire that ran from the 2018 charter through the 2025 “independent expert panel” mechanism ². The perverse incentive — OpenAI financially discouraged from declaring AGI, Microsoft incentivized to deny it — is replaced by a flat commercial calendar. Symbolically large, commercially overdue.

The financial story is more contested than vendor coverage admits

TD Cowen estimates the elimination of Microsoft’s revenue-share-out to OpenAI saves Redmond roughly $700M in FY26, scaling past $5B by 2030 ³. That tailwind explains Nadella’s public endorsement. But the buy-side dissent is sharper:

OpenAI commitments still drive roughly 45% of Microsoft’s AI backlog, posing a concentrated ‘major risk’ if the relationship continues to loosen ⁴.

Oppenheimer cut its MSFT target to $515; the stock traded down ~2% intraday. The real question isn’t whether the unwind helps margins — it’s whether MAI and Phi mature into a credible hedge before that 45% concentration unwinds faster than expected.

Bedrock parity is a claim, not a benchmark

AWS published GPT-5.5 at $5/$30 per 1M input/output tokens, GPT-5.4 at $2.50/$15, with Batch at 50% off and a Priority tier promising ~25% lower latency — and, critically, spend counts against existing AWS commits ⁵. That last point is the enterprise hook. Pricing parity with OpenAI’s direct API is the headline.

Practitioners aren’t buying parity on faith. Hacker News threads converged on three concrete worries about both GPT-5.4 and GPT-5.5 on Bedrock: quantization drift, inference-silicon differences (Trainium vs. Nvidia), and batching strategies producing subtle output deltas versus OpenAI’s native endpoint ⁶. The Claude-on-Bedrock precedent cuts both ways — better throughput, occasional quality regressions — and independent benchmark verification across the Bedrock, Azure, and native-OpenAI backends isn’t yet available.

What’s actually at stake

The amendment trades a mission-era romance for an IPO-era contract. Microsoft keeps the IP license through 2032 and a capped revenue stream; OpenAI gets multi-cloud distribution and the freedom to take Amazon’s money. Whether that’s a clean divorce or the first crack in the equity relationship depends on two unresolved questions: whether Bedrock’s GPT-5.4/5.5 actually behave identically to the native OpenAI endpoint under load, and whether Microsoft’s in-house models close the capability gap before the 45% backlog concentration starts to slip.

Further reading

• OpenAI models, Codex, and Managed Agents come to AWS — openai-blog
• OpenAI ends its exclusive partnership with Microsoft — ars-technica-ai
• Tracking the history of the now-deceased OpenAI Microsoft AGI clause — simon-willison

---

The Musk v. Altman trial is narrower than the headlines — and weirder

Source: mit-tech-review-ai · published 2026-04-27

TL;DR

• The jury is advisory only; Judge Gonzalez Rogers decides remedies and has already mocked Musk’s $134B damages number.
• Musk dropped fraud claims pre-trial; what’s left is equitable charitable-trust theory the state AGs declined to bring.
• OpenAI used opening day to end Microsoft’s cloud exclusivity — litigation theater against the “captured subsidiary” framing.
• Discovery surfaced a Sutskever plan to merge OpenAI into Anthropic during the 2023 board ouster.

What’s actually being tried

The MIT Tech Review and Ars Technica framing — that a Northern California courtroom could unwind OpenAI’s for-profit conversion and torpedo its IPO — is technically true and practically misleading. Musk strategically dropped his fraud claims ahead of trial, leaving only equitable claims of breach of charitable trust and unjust enrichment ⁷. That choice routes the case to a bench remedy: the nine-person jury is advisory only, and Judge Yvonne Gonzalez Rogers retains full authority over liability and any disgorgement ⁸. She has already signaled where she sits, telling the court she was “not particularly impressed” by the Musk expert’s $134 billion disgorgement calculation ⁸.

The standing problem is the part the primary coverage skips entirely. Under California and Delaware charitable-trust law, enforcement authority over a nonprofit’s purpose normally rests with the state Attorney General — and both AGs settled with OpenAI in October 2025 and declined to bring this case, leaving Musk to pursue a novel private-donor theory ⁹. OpenAI’s lead counsel William Savitt then made the standing fight personal in cross, surfacing 2017 emails in which Musk himself demanded 55% equity and majority control of a proposed for-profit OpenAI subsidiary — and framing the suit as “sour grapes” from a founder who lost an internal power struggle ¹⁰.

The maneuvering happening alongside the trial

The more interesting story is what OpenAI is doing outside the courtroom while the trial runs. On April 27, the same day opening statements began, OpenAI and Microsoft announced a “major amendment” ending Microsoft’s exclusive cloud relationship and clearing OpenAI to serve models on AWS and Google Cloud ¹¹. The timing is not subtle: the restructuring is widely read as a defensive move against Musk’s central claim that OpenAI has been “captured” as a Microsoft subsidiary in violation of its charitable purpose. Whether or not the judge buys that framing, the corporate facts on the ground have already shifted to undermine it.

Discovery has produced the trial’s other durable artifact. Internal records show Ilya Sutskever circulated a 52-page memo coordinated via “disappearing email” services with independent board members, and that the November 2023 plan included merging OpenAI with Anthropic immediately after firing Altman — with Helen Toner reportedly arguing destroying the company could be “consistent with the mission” ¹².

“Consistent with the mission” — Toner, on dismantling OpenAI

What’s actually at stake

Strip away the IPO-killing scenario and the trial’s likely contribution is evidentiary, not structural. A skeptical bench, an advisory jury, AG non-participation, and Musk’s own 2017 emails all cut against a sweeping remedy. But the Sutskever memo, the Anthropic merger plan, and the Microsoft cloud unwind are now part of the permanent public record regardless of how Gonzalez Rogers rules. The verdict may matter less than the discovery.

Further reading

• Musk and Altman face off in trial that will determine OpenAI’s future — ars-technica-ai

---

Anthropic’s creative-tools coalition lands to a Blender revolt

Source: anthropic-news · published 2026-04-28

TL;DR

• Anthropic shipped MCP-based Claude connectors for Adobe, Autodesk, Blender, Ableton, Splice, Affinity, SketchUp and Resolume.
• A €240,000/year Blender Development Fund patronage triggered immediate artist backlash; the Foundation is now “evaluating the situation.”
• Independent tests show Claude is competent on scaffolding and docs, unreliable on precision work like retopology or manufacturable CAD.
• The launch lands amid an unresolved copyright fight Anthropic is on both sides of.

A coalition built on MCP, greeted with a middle finger

Anthropic’s “Claude for Creative Work” launch is structurally ambitious: a single model acting as the orchestration layer across 50+ Adobe Creative Cloud tools, Autodesk Fusion, Blender’s Python API, Ableton Live, Splice’s sample library, Affinity, SketchUp and Resolume — all wired through the open Model Context Protocol so the connectors aren’t Claude-locked.

flowchart LR
    U[Artist/designer] --> C{Claude + Claude Code}
    C -->|MCP| A[Adobe CC ~50 tools]
    C -->|MCP| B[Blender Python API]
    C -->|MCP| F[Autodesk Fusion]
    C -->|MCP| M[Ableton / Splice / Resolume]
    C -->|MCP| S[SketchUp / Affinity]

The reception has been less tidy. Within hours of the Blender Foundation publicizing Anthropic’s €240,000-per-year patronage ¹³, the Foundation appended a notice to its own press release saying it was “actively evaluating the situation” in light of feedback ¹⁴ — an unusual hedge for a deal you just announced. New CEO Francesco Siddi conceded “there was not enough consideration” of how the donor would be perceived, while defending the move on the grounds that “participation does not imply alignment” ¹⁵. Artist Amos Mulder’s reaction circulated widely:

a middle finger to creatives… legitimizing a technology that threatens their livelihoods ¹⁶

The subtext inside the Blender community is generational: the deal arrived roughly four months into Siddi’s tenure after Ton Roosendaal handed over the CEO seat on Jan 1, and many longtime contributors read it as a policy break that wouldn’t have happened on the founder’s watch.

Competent junior, unreliable senior

Strip away the demo reel and Claude’s performance is uneven in ways that matter for production. SnapEDA’s February benchmark found Claude Opus 4.6 leading Gemini 3 at extracting mechanical dimensions from datasheets but still producing “occasional errors” ¹⁷ — fine for ideation, dangerous for parts that get manufactured. The Blender connector handles lighting setups, scene scaffolding and Python-API tutoring well, but stumbles on retopology and photogrammetry cleanup, the tasks that actually gate a professional pipeline. Ableton and Resolume integrations are strong as documentation oracles and for repetitive busywork, weaker on third-party VSTs and nested session state.

That maps cleanly onto Anthropic’s own five-pillar pitch: tutoring, scripting via Claude Code, pipeline glue, prototyping (the new Claude Design tool), and production automation. The first three and the last one are genuinely useful today. The middle — “rapid prototyping” of finished work — is where the marketing outruns the artifacts.

The copyright contradiction nobody resolved

The launch lands in legal quicksand. Anthropic settled a shadow-library copyright suit for $1.5B in late 2025, and a 2026 Claude Code leak prompted accusations that Anthropic was issuing DMCA takedowns to protect its own AI-assisted code while still arguing that ingesting artists’ work for training is fair use ¹⁸. The U.S. Supreme Court declined the same week to revisit human-authorship rules for AI output, leaving anything shipped through these connectors in an undefined zone of ownership.

The technical story here is real: MCP is becoming the de facto integration substrate for creative software, and Anthropic just wrote the reference implementation. The political story is that the open-source artist community Anthropic most needs for legitimacy is the one most visibly rejecting the deal — and the foundation taking the money is publicly second-guessing it.

Round-ups

GPT-5.5 ships with Codex folded into the main model line

Source: simon-willison

OpenAI is collapsing its Codex coding model into the flagship line, with Romain Huet confirming there will be no separate GPT-5.5-Codex release. The unified GPT-5.5 instead claims stronger agentic coding and computer-use gains, ending the parallel coding-specialist track introduced under GPT-4.

Further reading:

• [AINews] GPT 5.5 and OpenAI Codex Superapp — latent-space

OpenAI available at FedRAMP Moderate

Source: openai-blog

ChatGPT Enterprise and the OpenAI API received FedRAMP Moderate authorization, clearing both products for handling controlled unclassified information inside U.S. federal agencies and putting OpenAI on closer compliance footing with Microsoft and Google’s government cloud offerings.

EU tells Google to open up AI on Android; Google says that’s “unwarranted intervention”

Source: ars-technica-ai

The European Commission is pressing Google to let rival AI assistants reach parity with Gemini on Android under the Digital Markets Act. Google called the preliminary findings an ‘unwarranted intervention’ and is contesting the proposed remedies.

China kills Meta’s acquisition of Manus as US-China AI rivalry deepens

Source: ars-technica-ai

Chinese regulators blocked Meta’s planned acquisition of agent startup Manus, forcing the deal to unwind and underscoring how founders with mainland ties struggle to exit to U.S. buyers as Beijing tightens scrutiny over outbound AI talent and IP transfers.

[AINews] DeepSeek V4 Pro (1.6T-A49B) and Flash (284B-A13B), Base and Instruct — runnable on Huawei Ascend chips

Source: latent-space

DeepSeek released V4 in two MoE sizes — a 1.6T-parameter Pro with 49B active and a 284B Flash with 13B active — both shipping base and instruct checkpoints and tuned to run on Huawei Ascend silicon, though no longer topping benchmark leaderboards.

Speech translation in Google Meet is now rolling out to mobile devices

Source: simon-willison

Google Meet’s real-time speech translation, which mimics the speaker’s voice in the listener’s language with a short delay, is reaching mobile after a web rollout. It currently supports English, Spanish, French, German, Portuguese, and Italian, and Simon Willison reports mixed results across devices.

[AINews] ImageGen is on the Path to AGI

Source: latent-space

Latent Space’s newsletter argues image generation has become an AGI bellwether, citing the ongoing GPT-Image-2 surge as evidence that multimodal world-modeling — not text reasoning alone — is now driving frontier capability gains.

Footnotes

1. FinancialContent — FTC AI Interoperability ruling — https://www.financialcontent.com/article/marketminute-2026-4-8-breaking-the-walled-garden-ftc-mandates-ai-interoperability-for-tech-giants

   the U.S. Federal Trade Commission (FTC) took more aggressive action on April 8, 2026, by mandating ‘AI Interoperability’… requires large-cap providers like Microsoft and Google to allow enterprise customers to move fine-tuned models and prompts across competing clouds without financial penalties

   ↩

2. Simon Willison’s blog — https://simonwillison.net/2026/Apr/27/now-deceased-agi-clause/

   the new agreement describes revenue payments as ‘independent of OpenAI’s technology progress,’ effectively killing the AGI milestone as a legal trigger

   ↩

3. TD Cowen analyst note via Investing.com — https://www.investing.com/news/analyst-ratings/td-cowen-reiterates-buy-on-microsoft-stock-cites-openai-deal-93CH-4645469

   TD Cowen analysts estimate this shift will save Microsoft approximately $700 million in fiscal year 2026, potentially rising to over $5 billion by 2030

   ↩

4. TipRanks (analyst roundup) — https://www.tipranks.com/news/heres-how-analysts-reacted-to-microsofts-msft-loss-of-exclusivity-to-openai-technology

   OpenAI commitments still drive roughly 45% of Microsoft’s AI backlog, posing a concentrated ‘major risk’ if the relationship continues to loosen

   ↩

5. AWS Bedrock OpenAI page — https://aws.amazon.com/bedrock/openai/

   GPT-5.5: $5.00 per 1M input tokens / $30.00 per 1M output tokens… pricing aligns with OpenAI’s direct API rates, allowing customers to apply usage costs toward their existing AWS cloud commitments

   ↩

6. Hacker News discussion — https://news.ycombinator.com/item?id=47939320

   Developers expressed skepticism about whether GPT-5.4 or 5.5 on Bedrock would perform identically to the versions on OpenAI’s native platform or Azure… differences in quantization, custom inference silicon (like AWS’s Trainium), and batching strategies might lead to subtle degradations

   ↩

7. MLQ.ai on dismissed fraud claims — https://mlq.ai/news/us-court-dismisses-musks-fraud-allegations-against-openai-on-his-request-ahead-of-trial/

   Musk strategically dropped his fraud charges ahead of trial to streamline proceedings into equitable claims of breach of charitable trust and unjust enrichment — a bifurcated structure where the judge, not jury, fashions remedies.

   ↩ ↩²

8. GeekWire courtroom dispatch — https://www.geekwire.com/2026/musk-v-altman-inside-the-courthouse-as-microsofts-13-billion-openai-bet-goes-on-trial/

   Judge Gonzalez Rogers allowed Musk’s damages expert to testify but noted she was ‘not particularly impressed’ by his $134 billion disgorgement calculation; the nine-person jury serves only an advisory role, with the judge retaining ultimate authority on liability and remedies.

   ↩ ↩² ↩³

9. Duke Law / Daily Journal analysis — https://law.duke.edu/web/sites/default/files/pdf/Daily_Journal-Open_AI.pdf

   Under California and Delaware charitable trust law, the authority to enforce a nonprofit’s purpose typically rests with the state Attorney General, not individual donors — and the AGs, having already settled with OpenAI in October 2025, declined to bring this case, leaving Musk to pursue a novel legal path as a private donor.

   ↩ ↩²

10. Business Insider — https://www.businessinsider.com/elon-musk-sam-altman-openai-trial-testimony-2026-4

    OpenAI’s lead counsel William Savitt presented internal 2017 emails suggesting Musk himself had demanded majority control and 55% equity in a proposed for-profit subsidiary, framing the lawsuit as ‘sour grapes’ from a founder who failed to take over the company.

    ↩

11. Finimize — https://finimize.com/content/musk-takes-openai-to-court-over-its-for-profit-shift

    Coinciding with the start of the trial, Microsoft and OpenAI announced a ‘major amendment’ on April 27, 2026… ending Microsoft’s exclusive cloud partnership and allowing OpenAI to serve its models on AWS and Google Cloud — widely viewed as a strategic defense against Musk’s claim that OpenAI is a ‘captured’ subsidiary.

    ↩

12. MLQ.ai on Sutskever deposition — https://mlq.ai/news/ilya-sutskevers-secret-memo-and-deposition-revealed-in-musks-lawsuit-against-openai/

    Sutskever utilized ‘disappearing email’ services to coordinate with independent board members… and a previously undisclosed plan to merge OpenAI with its primary rival, Anthropic, immediately following Altman’s firing, with Helen Toner advocating that destroying the company could be ‘consistent with the mission.’

    ↩

13. CGChannel — https://www.cgchannel.com/2026/04/ai-developer-anthropic-becomes-blenders-latest-corporate-patron/

    Anthropic joins the Blender Development Fund as a Corporate Patron with a €240,000 annual contribution

    ↩

14. Blender Foundation (official notice) — https://www.blender.org/press/anthropic-joins-the-blender-development-fund-as-corporate-patron/

    We have received a lot of feedback regarding this announcement and are actively evaluating the situation.

    ↩

15. 80.lv — interview with Blender CEO Francesco Siddi — https://80.lv/articles/blender-ceo-on-anthropic-funding-this-is-not-ai-takeover

    in this case, it looks like there was not enough consideration… participation does not imply alignment

    ↩

16. Mastodon — Amos Mulder (artist) — https://mastodon.social/@amosmulder/116483663831928150

    a middle finger to creatives… legitimizing a technology that threatens their livelihoods

    ↩

17. SnapEDA benchmark blog — https://blog.snapeda.com/2026/02/16/evaluating-the-accuracy-of-ai-generated-cad-models/

    Claude Opus 4.6 outperformed competitors like Gemini 3 in extracting mechanical dimensions, though it still produced occasional errors

    ↩

18. TheFirewall blog — https://www.thefirewall-blog.com/2026/04/ai-code-leak-exposes-the-fault-lines-of-copyright/

    Anthropic drew criticism for using DMCA takedowns to protect its own AI-assisted code while simultaneously arguing that ingesting human artists’ work for training constitutes fair use

    ↩