# Coding agents picked sides. Compute drew the lines.

URL: https://www.thedeepfeed.ai/posts/2026-10-16-coding-agents-pick-sides/
Category: Business
Published: 2026-10-16
Author: the-deep-feed
Tags: cursor, spacex, openai, anthropic, coding-agents, claude-code, agents-md
Kind: deep

> OpenAI is pulling its models from SpaceX-owned Cursor while Anthropic, which rents $1.25 billion a month of SpaceX compute, promises Cursor more Claude, and the new X Subscription now prices every rival model as an extra.

## TL;DR

- On Aug 28, **OpenAI** gave SpaceX notice that it will stop serving its models to **Cursor** on a proposed **Nov 12** shutoff date and will send Cursor no future models, citing Musk companies' record on contracts and its next model, Astra.
- The same night, **Anthropic** co-founder Tom Brown promised Cursor more Claude capacity. Anthropic pays SpaceX **$1.25 billion a month** for Colossus compute through May 2029, which explains the warmth better than any partnership history.
- The sharper weapon is the bill. Since **Oct 12**, new Cursor customers buy **The X Subscription**, where Grok and Composer draw from an included pool and Claude, GPT and Gemini are billed at list price from prepaid credit.
- The harness layer is converging, not splitting: **Claude Code** began reading **AGENTS.md** on Sep 18 and Pi adopted MCP on Sep 30. Lock-in is moving below the editor, to compute, subscriptions and source control.

On the evening of Friday, August 28, OpenAI published a four-paragraph post that redrew the coding-agent market. [It had notified SpaceX](https://openai.com/index/our-decision-on-cursor-following-its-acquisition-by-spacex/) that it intended "to wind down our contract providing OpenAI models to Cursor, with a proposed shutoff date of November 12, 2026." The reason was stated without euphemism: "we cannot be confident that SpaceX will use our technology within our terms of service, based on our experience with Elon Musk's companies violating contracts." OpenAI added that it would hold the cancellation to the latest date the contract allowed "while not providing future models to Cursor." This publication [reported the notice](/posts/2026-08-30-openai-winds-down-cursor/) on August 30.

Cursor's chief executive answered that night. Anthropic answered 36 minutes after him.

> We're sorry to see that OpenAI put out a note saying they plan to block Cursor users from accessing OpenAI models in three months.
>
> OpenAI models serve about 5% of Cursor user traffic, and we're speaking with the OpenAI team to resolve this.
>
> Cursor was one of the very first users of OpenAI, we've worked closely with their team for years, and we've trusted their platform to be neutral infrastructure for our business.
>
> — [@mntruell](https://x.com/mntruell/status/2093532254006063557), Aug 28, 2026

> Cursor has been a trusted partner of Anthropic since Sonnet 3.5. We'll continue to increase compute to support Claude models in Cursor and are excited for what comes next with them at SpaceX.
>
> — [@NotTomBrown](https://x.com/NotTomBrown/status/2093541294027280657), Aug 28, 2026

The working hypothesis for this piece, going in, was simple: coding agents stopped being neutral tools, the editor, the model and the compute provider are consolidating into rival alliances, and model access is now a weapon. Seven weeks of filings, changelogs and pricing pages support half of that. The alliances are real, and one lab did use access as a weapon. But the lines do not run between labs the way a cutoff story implies. They run along compute contracts and subscription bills, and the weapon that does the most sorting in late 2026 is not the shutoff. It is the price sheet.

![Schematic: Anthropic pays SpaceX's Colossus $1.25B a month and supplies Claude to Cursor; OpenAI's link to Cursor is cut Nov 12](/post-images/2026-10-16-coding-agents-pick-sides/alliance-map.jpg)

# What SpaceX bought, in its own words

The premise needs checking first, because the HN headline that carried this story (852 points, [544 comments](https://news.ycombinator.com/item?id=49486172)) assumed it. SpaceX did acquire Cursor. Its [Form 8-K](https://www.sec.gov/Archives/edgar/data/1181412/000162828026056945/spcx-20260814.htm) records that a June 16 merger agreement with Anysphere, Inc. became effective on August 14, 2026, converting Cursor's stock into 389,289,254 shares of SpaceX Class A common stock "based on an implied equity value of Cursor of $60.0 billion." Cursor's own post that day, [Cursor is now a part of SpaceX](https://cursor.com/blog/joining-spacex), promised access to "the largest fleet of GPUs in the world" and "more capable models at lower cost."

The deal was not a surprise. In June, this publication [read xAI's Grok Build launch](/posts/2026-06-01-grok-build-and-the-model-xai-didnt-make/) against a $60 billion option SpaceX held on Cursor, struck in April alongside a compute agreement. What the August close added is a document that says, in SpaceX's own words, why it wanted the company. The [IPO prospectus](https://www.sec.gov/Archives/edgar/data/1181412/000162828026040364/spaceexplorationtechnologib.htm) describes the April deal as "a compelling extension of our strategy to vertically integrate compute infrastructure, models, and applications," and then names the asset:

> The depth of Cursor's integration with a high-frequency coding workflow generates valuable developer interaction data, including coding generation prompts, iteration cycles, and software architecture decisions. We expect that access to this data will enhance our model training and inference, including with respect to Grok.
>
> — SpaceX IPO prospectus, [Recent Developments: Collaboration with Cursor](https://www.sec.gov/Archives/edgar/data/1181412/000162828026040364/spaceexplorationtechnologib.htm)

That paragraph is the hinge of the whole OpenAI decision. A coding editor that routes a frontier model's output through its own interface sees every prompt, every rejected diff and every accepted one. When the editor's owner also trains a competing model and has said, in a filing, that it expects that interaction data to improve that model, the model vendor is no longer selling to a customer. It is selling to a training pipeline with a customer attached.

OpenAI's post makes the same argument from the other side. It cites Twitter breaking the terms of an earlier contract after Musk's takeover, and Musk's admission under oath this year that xAI had violated OpenAI's terms of service, conduct widely reported as distillation of OpenAI model outputs. It then names the asset it is protecting: "As AI capabilities advance, we also have a new level of accountability to ensure our upcoming model, Astra, is being used in accordance with our terms." Early readings on HN split between bad blood and model protection. One commenter [read the move](https://news.ycombinator.com/item?id=49486345) as protecting Astra "against distillation"; another [called it](https://news.ycombinator.com/item?id=49486319) "a pretty standard circling of the wagons in preparation for the next phase in the battle for frontier AI dominance."

# A cutoff aimed at 5% of the traffic

Measured by what it removes, OpenAI's weapon is small. Michael Truell's own figure is that OpenAI models served about 5% of Cursor user traffic. The notice uses the maximum period the contract allows, and Cursor kept serving OpenAI models through it: at the time of writing, [Cursor's model page](https://cursor.com/docs/models) lists GPT-5.6 Luna, GPT-5.6 Sol and GPT-5.6 Terra. What it does not list is any model from the GPT-6 generation OpenAI has shipped since, including the [GPT-6.1 Sol tiers whose pricing this publication covered earlier this month](/posts/2026-10-09-gpt-6-1-sol-prices-speed-above-astra/). The "no future models" clause has already taken effect. The shutoff itself has not.

Musk's reply, a few hours later, was [a post on X](https://x.com/elonmusk/status/2093572368434127259) that began "I couldn't care less," followed by insults aimed at Sam Altman and Greg Brockman. The response reads as temperament, and it also reads as arithmetic. A coding product that already ran 95% of its traffic on other vendors' models and its own can absorb the loss of one vendor's catalog, particularly when its parent is building a replacement on [the largest GPU fleet in the world](https://cursor.com/blog/joining-spacex). [CNBC's account](https://www.cnbc.com/2026/08/29/openai-cursor-spacex-model-access.html) of the exchange puts the dispute inside the longer Musk and OpenAI fight, including a lawsuit Musk lost earlier this year and has vowed to appeal.

The more useful question is what the cutoff tells other editors. Every independent coding product that resells frontier models now knows that a change of control can trigger a vendor's exit clause, and that at least one lab will use it. OpenAI's post confirms the mechanism: "Our custom agreement with Cursor gives us a limited time window to cancel it after a change of control." Any coding startup courting an acquirer that also builds models now carries that clause into the negotiation as a discount.

![Timeline from SpaceX's April option on Cursor through the Aug 14 close and Aug 28 notice to the proposed Nov 12 shutoff](/post-images/2026-10-16-coding-agents-pick-sides/cutoff-timeline.jpg)

# The partner that could not afford to leave

If model access were simply a weapon, Anthropic would have been the first to fire it. It has the clearest record of doing so. In June 2025, Windsurf said Anthropic had cut its direct access to Claude models while reports circulated that OpenAI would buy it; Anthropic co-founder Jared Kaplan told a TechCrunch event that ["it would be odd for us to be selling Claude to OpenAI."](https://techcrunch.com/2025/06/05/anthropic-co-founder-on-cutting-access-to-windsurf-it-would-be-odd-for-us-to-sell-claude-to-openai/) In January of this year, according to [The New Stack's reconstruction](https://thenewstack.io/anthropic-spacex-cursor-compute-openai/), Anthropic cut xAI staff off from Claude they had been reaching through Cursor, with Cursor reportedly describing it as a policy Anthropic was enforcing against all its major competitors.

This time Anthropic did the opposite, and two of the sharpest replies to Tom Brown's post said why.

> @NotTomBrown Maybe you changed your ways, but we all remember what you did to Windsurf, which was infinitely nastier.  More likely answer is that you can't do that here because you need the compute.
>
> — [@amasad](https://x.com/amasad/status/2093543108491530308), Aug 28, 2026

> Either SpaceX and Grok are not competitors to Anthropic (they are!); or, more likely, SpaceX leasing its Colossus 1 data center is more important to Anthropic than to stop offering Claude to SpaceX.
>
> SpaceX basically in this enviable position where one of its biggest competitors depends on its compute infra!
>
> — [@GergelyOrosz](https://x.com/GergelyOrosz/status/2094006046138282122), Aug 30, 2026

The filings back the critics. On May 6, Anthropic [announced it would use all of the compute at SpaceX's Colossus 1](https://www.anthropic.com/news/higher-limits-spacex), "more than 300 megawatts of new capacity (over 220,000 NVIDIA GPUs) within the month," and used the capacity to double Claude Code's five-hour rate limits on the same day. SpaceX's prospectus supplies the terms Anthropic's post left out: cloud services agreements covering "approximately 325,000 NVIDIA GPUs" across Colossus and Colossus II, under which Anthropic "has agreed to pay us $1.25 billion per month through May 2029." After an initial three-month period, either side may terminate on 90 days' notice, and the prospectus notes the structure lets SpaceX reallocate that capacity "for our own internal initiatives if needed in the future."

That clause is what makes the relationship lopsided. Anthropic is paying a direct competitor $15 billion a year for capacity that backs its own subscribers' usage limits, under a contract the competitor can end with a quarter's warning. Cutting off Claude inside the competitor's coding product would be principled, and it would also test the patience of a landlord. Anthropic chose to deepen the tie. On September 9, the Claude account announced Cursor as one of five new vendors in Claude Marketplace, where enterprise customers can spend their Anthropic commitments on partner products.

> New on Claude Marketplace: @CrowdStrike, @cursor_ai, @FactoryAI, @GammaApp, and @vercel.
>
> Enterprises can now use their Anthropic spend commitment to buy more Claude-powered products and agents.
>
> — [@claudeai](https://x.com/claudeai/status/2097718980437831935), Sep 9, 2026

So the alliance map has a structural oddity at its center. OpenAI, which owes SpaceX nothing, can afford to cut. Anthropic, which owes SpaceX $1.25 billion every month, cannot, and it now sells a SpaceX subsidiary's product through its own enterprise channel. In August this publication argued that [Stripe's reported OpenRouter deal bought the meter on inference demand](/posts/2026-08-18-stripe-bought-the-meter/). The Colossus contract is the same lesson from the supply side: whoever owns the GPUs can collect rent from a rival and still be treated as a partner by it.

# The bundle does the sorting

Model access, then, turned out to be a weapon only one lab could afford to use, and it hit 5% of one product's traffic. The weapon that reaches every Cursor customer arrived on October 12, and it was a pricing page.

That day, Cursor [introduced The X Subscription](https://forum.cursor.com/t/introducing-the-x-subscription/174450), which "brings together Grok, Grok Bot, Cursor, and X Premium in one plan." New customers start on it, and existing Cursor customers can keep their plans until they switch. The structural change sits in one line of the [legacy-plans page](https://cursor.com/docs/account/pricing/legacy-cursor-plans): each X Subscription tier "includes at least twice as much coding usage with Cursor models as the legacy plan it replaces," and "it also drops the included third-party model allowance; third-party models are billed separately."

| X Subscription plan | Price | Included pool (Grok, Composer) | Claude, GPT, Gemini |
|---|---|---|---|
| Plus | $8/mo | Base | Not available |
| Premium | $30/mo | 4x Plus | API list price, prepaid credit |
| Super | $100/mo | 15x Plus | API list price, prepaid credit |
| Ultra | $200/mo | 50x Plus | API list price, prepaid credit |
| Teams | $30 to $200 per seat | Matches seat tier | List price plus $0.25 per million tokens |

Under the old plans, a Cursor Pro subscriber had two monthly pools, one for Cursor's own models and one called Other Models that paid for Claude, GPT and Gemini at their API prices. Under the new plans, the included pool covers Grok 4.7, Grok 4.6, Grok 4.5 and Composer 2.5, refreshes weekly, and is shared with the Grok apps. A third-party model draws nothing from it. On the [models page](https://cursor.com/docs/models), the cheapest paid tier, Plus at $8, excludes third-party models and bring-your-own API keys entirely. Teams and Enterprise customers pay a Cursor Token Rate of $0.25 per million tokens on every third-party request, "on top of model API pricing for additional usage and BYOK usage," while "Cursor models, including Grok and Composer" are exempt.

None of this blocks a rival model. It changes the default. An individual developer on Premium who never thinks about vendors will spend the included pool on Grok and Composer, because that usage is already paid for. Choosing Claude Opus 5.5 now means choosing to spend extra prepaid credit at $4 per million input tokens and $20 per million output. When Cursor Router picks a model for a team, a route to a third-party model bills as additional usage. Weeks before the bundle launched, an HN commenter [had already described the economics](https://news.ycombinator.com/item?id=49486323): "Cursor is already kind of useless for third party models unless you're willing to spend thousands of dollars. It's only worth it if you're going to use mostly grok/composer."

![Billing schematic: a weekly included pool for Grok and Composer, a prepaid-credit pipe with a toll for rival models](/post-images/2026-10-16-coding-agents-pick-sides/bundle-billing-flow.jpg)

The shelf, meanwhile, stays open. Cursor added Google's newest model the day after the bundle launched, and Meta's two days after that.

> Gemini 4 Argon is now available in Cursor! It scores 52.8% on CursorBench at default effort, on par with Opus 5.5. Priced at $2/M input and $10/M output tokens.
>
> — [@cursor_ai](https://x.com/cursor_ai/status/2110087611763368271), Oct 13, 2026

> Muse Wonder from Meta Superintelligence Labs is now available in Cursor! It scores 46.1% on CursorBench at default effort, the highest of any Muse model. Priced at $4/M input and $20/M output tokens.
>
> — [@cursor_ai](https://x.com/cursor_ai/status/2110867556693430293), Oct 15, 2026

Those posts are the clearest picture of what an alliance looks like in this market. It is not a closed store. It is a store with a house brand on the free shelf and every other brand at full retail, with a toll at the register for business buyers. Supermarkets have run that model for decades. What is new is that the house brand's owner also writes, in a securities filing, that it expects the shoppers' behavior to train the house brand.

# Employers are picking sides too

The same sorting is happening inside the companies that build models, and there the instrument is the internal budget rather than the price sheet. According to The Information's October 5 report, as summarized by [Channel Insider](https://www.channelinsider.com/ai/news-microsoft-meta-cut-internal-claude-use/), Microsoft cut its projected internal spending on Anthropic technology, previously expected to exceed $1 billion a year, by more than a third, and lowered most individual monthly AI spending caps in its Cloud and AI division from $100,000 to roughly $10,000 while steering developers to GitHub Copilot. Internal Claude Code use at Meta reportedly fell from about 60,000 employees earlier this year to about 30,000, while its in-house MetaCode passed 30,000 internal users and Muse Code, the external Claude Code competitor, passed 6,000. Neither company has confirmed the figures, and both still sell Claude to customers through their platforms.

The HN thread on the report (378 points) [carried the cynical reading](https://news.ycombinator.com/item?id=49997575) that this is "about frontier AI shops dogfooding their own models," and another commenter [put it more bluntly](https://news.ycombinator.com/item?id=49997485): "Every big lab blocks competitor tools for internal use; it is a data governance thing, not a quality statement." Both readings fit the SpaceX prospectus. Developer interaction data is training data, and a lab that lets 60,000 engineers send their working sessions to a rival is, in effect, donating to that rival's next model.

Price does the rest of the sorting outside the labs. The FT reported on August 23 that Anthropic's most capable model was struggling to win users while cheaper tools thrived; [Simon Willison's notes on the story](https://simonwillison.net/2026/Aug/23/anthropics-best-ai-model-struggles-to-attract-users-as-cheaper-t/) reproduce Ramp's July breakdown, in which Claude Fable 5 took 8.0% of Anthropic model spend against 28.0% for Opus 4.8. [Analytics India Magazine](https://analyticsindiamag.com/ai-features/almost-nobody-is-using-anthropics-fable-5), also citing Ramp, put Fable 5 at 6% of the tokens businesses bought from Anthropic in its first two months, at an API price of $10 per million input tokens and $50 per million output. Buyers pick the model the bill favors. Cursor's new plans make Grok and Composer the model the bill favors.

# The files that make switching cheap

Underneath the alliances, the layer that developers touch every day is moving the other way. The instruction files, skills and tool protocols that tell a coding agent how a repository works are converging on shared formats, and the holdouts gave up during this window.

On September 18, Claude Code 2.1.277 [shipped this line](https://code.claude.com/docs/en/changelog): "Added AGENTS.md support: in a project with no CLAUDE.md, Claude Code reads AGENTS.md instead." AGENTS.md is the cross-vendor instruction file that OpenAI's Codex reads, and that xAI's Grok Build [advertised reading out of the box](/posts/2026-06-01-grok-build-and-the-model-xai-didnt-make/) in May. The HN post of the changelog drew 741 points. One commenter's [entire reaction](https://news.ycombinator.com/item?id=49760282) was "Vendor locks over md file names is crazy." Another [summed up the relief](https://news.ycombinator.com/item?id=49759925): "Finally. Now I can remove all my symlinks."

The support arrived reluctantly. Five days later, a developer [documented](https://blog.szypowi.cz/p/claude-code-reads-agents.md-only-when-telemetry-is-on/) that the loader shipped as a built-in plugin whose availability depended on a remote feature flag, so anyone who had disabled telemetry or non-essential traffic, along with teams on Bedrock, Vertex and gateways, silently never loaded the file. "Reading a markdown file from the working directory needs no network at all, but here it waits on a server-side switch," the post observed. Claude Code 2.1.281, released September 23, changed AGENTS.md support "to also work on Amazon Bedrock, Google Vertex AI, Microsoft Foundry, LLM gateways, and sessions with telemetry disabled." The gap between announcement and working feature was five days, and the episode shows where the vendor's instinct lay. It also left shared skills unsolved: the same post found Claude Code still ignores the `.agents/skills` directory that Codex reads.

The same week, the smaller harnesses moved in the same direction. On September 30, Earendil [explained](https://earendil.com/posts/you-said-no-mcp/) why Pi, a coding harness whose site once proudly declared it did not support MCP, now ships MCP in its core. [the post says](https://earendil.com/posts/you-said-no-mcp/) the team decided ["the best way to positively influence something is to embrace it,"](https://earendil.com/posts/you-said-no-mcp/) and it pitches a JavaScript sandbox it calls Codemode for composing MCP tool calls, "like other harnesses like Codex do too." The HN thread (687 points) was skeptical of the design and accepting of the outcome. One reply [compared MCP to USB-C](https://news.ycombinator.com/item?id=49906865): suboptimal, but "widely compatible and easy for the end user. That's why MCP is everywhere."

![Schematic: one AGENTS.md file and an MCP server feeding Codex, Claude Code, Grok Build and Pi, with a flag gate on one inlet](/post-images/2026-10-16-coding-agents-pick-sides/instruction-file-convergence.jpg)

Convergence at this layer is the reason the alliances have to fight elsewhere. When a repository's instructions, tool connections and skills port between harnesses with a symlink, the editor stops being a moat. A developer who leaves Cursor for Codex or Claude Code brings the AGENTS.md file along. What does not port is a prepaid weekly pool, an enterprise spend commitment that can only be spent in one marketplace, or a monorepo that only one company's infrastructure can serve.

# Where the lock-in moved: below the editor

The clearest view of that lower layer came from Anthropic itself on October 13. Its engineering post, [Redesigning source control at Anthropic](https://www.anthropic.com/engineering/redesigning-source-control) (890 points on HN), opens with the claim that "coding is no longer the bottleneck to building systems or doing research." The bottleneck is infrastructure built for human scale. Anthropic says it merged about 2,500 code changes a day in May and more than 30,000 a day by September. Its main branch history is almost 100 GB, it performs on the order of 25 million checkouts a day for CI, and fetching those naively "would be over 600 PB of git traffic per day."

The response was Scion, an in-house version control system built around the observation that "nearly all of our clients were agents and jobs that are online by definition." Clients fetch objects on demand from tiered caches; the server computes history, blame, merges and rebases; working copies are optional and sparse. Anthropic reports that its Scion footprint is 200 times smaller than the git mirroring infrastructure it replaced, that checkouts dropped from minutes to seconds, and that CI alone gained 25% of CPU capacity. A team of seven built Scion, a code review tool called Antique and related systems, and migrated the monorepo, in five months.

![Scion's tiers: agents and CI jobs fetch objects on demand through tiered caches from a server that computes history and merges](/post-images/2026-10-16-coding-agents-pick-sides/scion-architecture.jpg)

Two passages matter for the alliance question. The first is a disclaimer: "Scion is by no means a suitable replacement for git for general use," because it requires its own code review, code search and CI. Anthropic proposes instead to share its thinking so that "GitHub, GitLab, and the OSS community" can bring the useful parts to git. The second is about the people: "Very often, we found that the role of the engineers was to tell Claude 'no,'" because agents "will sometimes violate core system constraints to satisfy a pattern that should never have been allowed in the first place." The developer culture had already turned that complaint into comedy. A September interactive parody, [Opusfived](https://opusfived.dev/), asks the player to make one "Add to Cart" button blue and "not let Claude change anything else"; it reached 1,215 points on HN.

Cursor is building the commercial version of the same layer. Origin, its code host, [opened an app marketplace](https://cursor.com/changelog/origin-app-marketplace) on October 13 with more than 20 integrations for deployment, CI, code review, security, monitoring and project management, and opened it to public submissions. Origin is available on Premium and above. Anthropic keeps its forge internal and asks the git ecosystem to catch up; SpaceX sells its forge inside the subscription bundle. Either way, the contested ground is the system that stores, checks out and merges what the agents write, a layer that converges far more slowly than an instruction file does.

# November 12 tests the weakest weapon

The cutoff that made the headlines will be the least consequential event on this map. If OpenAI holds to its proposed date, Cursor loses three GPT-5.6 models that serve a small share of its traffic, and its customers lose a choice they increasingly paid extra to make. The stronger forces are already in place. SpaceX owns the editor, sets the price of every rival model inside it, and holds a compute contract its largest model-vendor tenant cannot easily replace. Anthropic sells through that editor because it needs the landlord's GPUs. Microsoft and Meta are steering their own engineers away from a rival's tools because a working session is training data. And the instruction files that once tied a developer to one harness now port with a symlink.

Coding agents did pick sides between August and October. The sides are drawn by who owns the compute, who issues the bill and who keeps the interaction data, and those lines cut straight through the partnerships the labs announce. The useful question for a team choosing a coding agent this quarter is no longer which model the tool supports. It is who owns the GPUs underneath it, which models its included usage pays for, and which company's training run its sessions will feed.

## Sources

- [OpenAI — Our decision on Cursor following its acquisition by SpaceX (Aug 28, 2026)](https://openai.com/index/our-decision-on-cursor-following-its-acquisition-by-spacex/)
- [SpaceX — Form 8-K, completion of the Anysphere (Cursor) merger (Aug 14, 2026)](https://www.sec.gov/Archives/edgar/data/1181412/000162828026056945/spcx-20260814.htm)
- [SpaceX — IPO prospectus: Cursor option agreement and Anthropic cloud services agreements](https://www.sec.gov/Archives/edgar/data/1181412/000162828026040364/spaceexplorationtechnologib.htm)
- [Cursor — Cursor is now a part of SpaceX (Aug 14, 2026)](https://cursor.com/blog/joining-spacex)
- [Anthropic — Higher usage limits for Claude and a compute deal with SpaceX (May 6, 2026)](https://www.anthropic.com/news/higher-limits-spacex)
- [Cursor Docs — Models and pricing (The X Subscription, Cursor Token Rate)](https://cursor.com/docs/models)
- [Cursor Docs — Legacy Cursor plans](https://cursor.com/docs/account/pricing/legacy-cursor-plans)
- [Cursor Forum — Introducing The X Subscription (Oct 12, 2026)](https://forum.cursor.com/t/introducing-the-x-subscription/174450)
- [Claude Code — Changelog (2.1.277, Sep 18, 2026; 2.1.281, Sep 23, 2026)](https://code.claude.com/docs/en/changelog)
- [Anthropic Engineering — Redesigning source control at Anthropic (Oct 13, 2026)](https://www.anthropic.com/engineering/redesigning-source-control)
- [Earendil — You said no MCP (Sep 30, 2026)](https://earendil.com/posts/you-said-no-mcp/)
- [CNBC — OpenAI to end model access to Cursor after acquisition by SpaceX (Aug 29, 2026)](https://www.cnbc.com/2026/08/29/openai-cursor-spacex-model-access.html)
- [The New Stack — Why Anthropic is sticking with Cursor as OpenAI cuts access](https://thenewstack.io/anthropic-spacex-cursor-compute-openai/)
- [Channel Insider — Microsoft, Meta cut internal Claude use as in-house AI tools take priority](https://www.channelinsider.com/ai/news-microsoft-meta-cut-internal-claude-use/)
- [Simon Willison — notes on the FT's Anthropic's best AI model struggles to attract users (Aug 23, 2026)](https://simonwillison.net/2026/Aug/23/anthropics-best-ai-model-struggles-to-attract-users-as-cheaper-t/)
- [Analytics India Magazine — Almost nobody is using Anthropic's Fable 5 (Aug 18, 2026)](https://analyticsindiamag.com/ai-features/almost-nobody-is-using-anthropics-fable-5)
- [Szypowicz — Claude Code reads AGENTS.md only when telemetry is on (Sep 23, 2026)](https://blog.szypowi.cz/p/claude-code-reads-agents.md-only-when-telemetry-is-on/)
- [TechCrunch — Anthropic co-founder on cutting access to Windsurf (Jun 5, 2025)](https://techcrunch.com/2025/06/05/anthropic-co-founder-on-cutting-access-to-windsurf-it-would-be-odd-for-us-to-sell-claude-to-openai/)
- [Cursor — Origin App Marketplace (changelog, Oct 13, 2026)](https://cursor.com/changelog/origin-app-marketplace)
- [Hacker News — Our decision on Cursor following its acquisition by SpaceX (852 points)](https://news.ycombinator.com/item?id=49486172)
- [@mntruell — Cursor CEO on the OpenAI notice (Aug 28, 2026)](https://x.com/mntruell/status/2093532254006063557)
- [@NotTomBrown — Anthropic co-founder on Claude in Cursor (Aug 28, 2026)](https://x.com/NotTomBrown/status/2093541294027280657)

---

Canonical: https://www.thedeepfeed.ai/posts/2026-10-16-coding-agents-pick-sides/
Site: https://www.thedeepfeed.ai
Full corpus: https://www.thedeepfeed.ai/llms-full.txt