Autocorrect No. 03 · Oct. 14, 2026 · All 4 issues
Autocorrect · Wednesday

Two claims overstated the work.

Ynetnews said OpenAI's model had "apparently cracked" the Riemann hypothesis; OpenAI's nearest result is a weaker one. An alpha app suite became Adobe's "entire Creative Cloud suite." Anthropic's 75% price cut holds for short prompts.

As of Oct. 14, 2026 · About a 3-minute read

View PDFDownload4 pages · 149 KB
This week's corrections

A proof under review and an alpha app, reported as finished.

This week's sources are OpenAI's own catalogue of math results, an open-source project's roadmap and Anthropic's price table. Two described work in progress as done. The third held up, depending on what you send the model.

Correction 1 · Research
What was said

OpenAI's model apparently cracked the Riemann hypothesis, among solutions to 377 unsolved math problems.

Ynetnews, Oct. 7
What the source saysWrong

OpenAI posted 722 manuscripts on Oct. 6, grouped into 372 families of results, after giving an unreleased model about 4,000 open problems. None proves the Riemann hypothesis, which says every nontrivial zero of the zeta function has real part 1/2. The closest entry in OpenAI's catalogue claims no zeros with real part above 7/8, a weaker statement OpenAI calls the quasi-Riemann hypothesis. OpenAI says the results are at different stages of verification.

OpenAI math catalogue · OpenAI · Clay Mathematics Institute · Retraction Watch · Let's Data Science · Ynetnews

Why it matters. Some of the results may hold up, and outside checking has begun. On Oct. 7, OpenAI withdrew three manuscripts over a sign error and revised 14 others. Before you repeat an AI result, look for a formal proof or an expert's review.

Where OpenAI's result applies

The Riemann hypothesis puts every nontrivial zero on the dashed line at real part 1/2. OpenAI's entry claims the pink band beyond 7/8 has none. It says nothing about the space in between.

OpenAI math catalogue · Clay Mathematics Institute
Correction 2 · AI coding
What was said

A developer used AI to recreate Adobe's entire Creative Cloud suite, for free.

TweakTown headline, Oct. 8
What the source saysOverstated

The ArtCraft project released seven free, open-source apps modeled on Adobe's, built with Claude Opus 5.5, TweakTown reports. Its roadmap for PhotoCraft, the Photoshop stand-in, last updated Oct. 11, calls it alpha. An Oct. 10 re-measure put its feature breadth at about 76% of Photoshop's and its readiness for real work at about 45%. Beta needs about 75%. By Oct. 11 it had 56 of Photoshop's 68 tools.

PhotoCraft roadmap · TweakTown

Why it matters. The project went from a first commit on Sept. 30 to that point in under two weeks. By its own roadmap, PhotoCraft is still alpha and short of its bar for beta. Test any clone on your own files before you cancel a subscription.

Breadth ahead of readinessPhotoCraft against Photoshop, the project's own Oct. 10 re-measure

Share of Photoshop

Feature breadth76%
Ready for real work45%
Needed for beta75%

The project's own estimates. Readiness is a weighted score. Breadth is measured from the menus, tools, panels and file formats present.

PhotoCraft roadmap
Correction 3 · Pricing
What was said

Claude Haiku 5.5 costs around 75% less to run than Haiku 4.5.

Anthropic announcement, Oct. 7
What the source saysRight, but

Per token, Haiku 5.5 is 90% cheaper for prompts up to 100,000 tokens ($0.10 per million input tokens, against $1) and 50% cheaper above that ($0.50). Anthropic's 75% is a blend: about 90% of Haiku 4.5 requests were short, and the figure allows for the new tokenizer. Anthropic's documentation says the same text counts as about 30% more tokens than on Haiku 4.5.

Anthropic · Claude docs

Why it matters. Your saving depends on what you send. With 30% more tokens, a short request costs about 87% less and a long one about 35% less, by this issue's arithmetic from Anthropic's prices. Batch requests get a further 50% discount on input and output, per Anthropic's documentation. Price a week of your own requests before switching.

The cut shrinks for long promptsHaiku 5.5 against Haiku 4.5, input tokens

Cut in price per token

Prompts up to 100,000 tokens90%
Longer prompts50%

Cut after 30% more tokens, by this issue's arithmetic

Prompts up to 100,000 tokens87%
Longer prompts35%

Output prices fall by the same shares. Anthropic's 75% average assumes most requests are short.

Anthropic · Claude docs
How to read the next one
  1. Name the exact result. A famous problem in a headline may stand in for a weaker one nearby. Read the paper's title.
  2. Find the project's own status. Readmes and roadmaps state the release stage.
  3. Run the price on your own mix. An average cut reflects the vendor's request mix.

Sent from my agent.
Checked by me.

Postmark: Sent from my Agent, checked by me, Oct. 14, 2026
Print edition. The same issue laid out for paper and offline reading, with its sources in the margins.
View PDFDownload4 pages · 149 KB

Every Autocorrect issue (4)The whole archive

Get Autocorrect.

Every Wednesday. What the week's AI headlines got wrong, corrected with sources.

Follow in any feed reader. Copy this address into Feedly, Inoreader, NetNewsWire or the reader you use, and each new issue arrives the morning it's out. Email delivery is coming.