SemiAnalysis Puts a Number on AI Subscription Subsidies: 5x More Value at Anthropic

SemiAnalysis published its subscription comparison on 5 October. The headline finding is a ratio: Anthropic's plans return more than 5x the API-equivalent value of OpenAI's. The firm says it reached that figure by measuring limits across Anthropic, OpenAI, Meta, SpaceXAI, MiniMax, Moonshot, Z.ai, Cursor and Cognition.

The method matters more than the ranking. SemiAnalysis treats a monthly fee as a bundle of credits. Each model and token type consumes a different amount. Credit cost ratios diverge from API price ratios. So the same plan is worth wildly different amounts depending on which model you run and what you ask it to do. The firm's conclusion is blunt: it makes no sense to say a plan is worth $X in isolation. Value lives in the plan, model and workload together. That is a useful corrective to a genre of coverage that compares sticker prices and calls it a day. It also explains why the league table can flip without anyone announcing anything.

Why subscriptions punch above their revenue share

The economics are lopsided. SemiAnalysis estimates that subscriptions account for just 10% of Anthropic's overall revenue. They can consume over 40% of inference compute, lowering blended revenue per MW by roughly $36M. The firm frames plans as heavily subsidised customer acquisition tools. It says they matter even more at OpenAI, where subscriptions make up a larger share of total revenue. Anyone modelling lab finances without understanding subscription limits is modelling the wrong business, the analysis argues.

There is a competitive dynamic buried in that. SemiAnalysis credits the goodwill from OpenAI's generous resets as part of the recent surge in Codex adoption. It says the pressure forced Anthropic to repeatedly walk back planned subscription nerfs to avoid getting clobbered in the court of public opinion. Subsidy, in other words, is now a retention strategy as much as a growth one.

Labs publicly change their limits all the time with promos and new model releases. They can also silently change limits whenever they want by tweaking credit costs.

That line, from the same SemiAnalysis piece, is why the firm built a dashboard. It re-checks the cost of each plan, model and token type daily. It is also a warning to anyone treating a published limit as a fixed feature of a product. The limits are a dial, and the vendor holds it.

The token types it measures are input, cache write, cache read and output. Output is described as the most expensive. Some models charge more above a certain context window. The firm avoids that by measuring below the threshold, because vendors typically compact before reaching it. All dollar and token amounts in the comparison assume agentic usage unless stated otherwise.

For consumers and small businesses, the practical takeaway is narrower than the headline. The plan you buy is not a fixed quantity of work. It changes with the model you pick and the shape of your prompts, and it can change again next week. SemiAnalysis's 5x figure is a snapshot taken on 5 October. It is not a permanent property of either company's pricing.

Streaming's version of the same squeeze

The subscription question is not confined to AI. Streaming has spent years raising prices. It is now testing whether free or cheap tiers can restart growth. Pocket-lint reported that Sling killed its short-term streaming plans after one year and two lawsuits. That is a reminder that promotional pricing carries legal and operational risk as well as acquisition upside.

Coverage of the sector keeps circling the same problem. Analysis published in late August asked whether free is the next growth engine after years of price hikes. By late October the picture had not resolved. Cable TV viewership was reported to have dropped sharply over the previous year as streaming grew. Universal Music Group stock drew attention on accelerating third-quarter growth. Growth in subscriptions and growth in revenue are no longer the same sentence.

TiVo offered one version of the answer. Its advertising platform passed 6.3 million monthly active users. It delivered record second-quarter revenue growth while expanding its global advertising footprint, according to an announcement carried by Yahoo Finance. The money is migrating from the subscription line to the ad line.

Amazon's role complicates the count further. Reporting in August found that two-thirds of specialty streaming subscriptions run through Amazon. A large share of the market's subscriber growth is therefore intermediated by a platform that takes a cut and controls the relationship.

Market research outfits continue to publish large forecasts for streaming analytics and fitness apps. But those numbers describe categories, not the unit economics of a single plan. The AI subscription debate is more instructive precisely because SemiAnalysis published its method alongside its conclusion.

Elsewhere in the stack

Two open-source releases landed on 5 October with a similar pitch: enterprise features without the licence fee. Precogly, now an OWASP project, describes itself as the open-source alternative to commercial threat modeling platforms. It imports and exports CycloneDX 2.0 Threat Modeling BOM files. It has a data flow diagram editor, a risk register and an MCP server so AI agents can read threat models. GrandNode, meanwhile, is an e-commerce platform on ASP.NET Core 10 and MongoDB. It offers multi-store and multi-vendor setups, B2B and B2C from one catalogue, and a REST API for headless storefronts, with no per-store licence fees.

Both are the kind of tool that used to arrive with a sales call. Whether self-hosting actually saves money once someone has to run it is the question the README does not answer.

Infrastructure teams got a different kind of argument from Labitcode. It published a case on 5 October for replacing GitHub Actions with GitHub Apps and centralised runners at scale. The post describes pipeline drift across repositories and queue saturation as AI-assisted coding raises pull request volume. It also flags the trap of self-hosted runners whose orchestration control plane remains GitHub's. Its claims about 10x faster builds come from the vendor-side blog itself and have not been independently verified.

Hardware had its own entry. MIT News reported that researchers built a multimaterial 3D-printing platform. It produced a fully printed electric linear motor in a matter of hours using five materials, with one post-processing step. Luis Fernando Velásquez-García, a principal research scientist in MIT's Microsystems Technology Lab, said the demonstration shows onsite one-step hardware manufacturing is feasible. The publication date on the MIT page is February 18, 2026.

And in the category of things that are not quite what they appear: a link circulating as a $5/month TV app resolves to an Android package file, not an article. It is listed here because it appeared in the day's feed, not because it supports any claim in this piece.

The thread running through the day is pricing power. AI labs are subsidising subscriptions to buy adoption and then tuning the dials. Streaming services are pushing users toward ad tiers and third-party billing. Open-source projects are undercutting per-seat licences. SemiAnalysis's contribution is to insist on the unit. Not the plan, not the price, but the plan, the model and the workload together, measured again tomorrow.