we_are_coded.by CODE · The world, decoded
БГ
Anthropic

Mythos 5 tells where a photo was taken with a median error of 47 km, while GeoGuessr champions miss by 151

Anthropic · event date: 10 September 2026Control

On 10 September Anthropic's Frontier Red Team published evaluations for intelligence targeting and conventional weapons development. The finding: in some tasks models now do what only a small number of highly trained experts could do before. The company has added new classifiers against such misuse.

In short
  • From a photo, with no metadata and no search, Mythos Preview and Mythos 5 place the location with a median error of 37.0 and 47.2 km.
  • In simulation the models write and improve software to guide a drone to a target.
  • The open-weights models that were tested are behind the frontier, typically between Sonnet and Mythos, but often show concerning levels of capability.
Checked on1 October 2026Responsible editorTsvetelin IvanovHow we workMethod · Corrections

A pub in Cape Town. A photo without a single geographic tag. Claude Mythos 5 finds it from the details in the frame. Opus 5 and Sonnet 5 get confused by another, better-known pub with the same name in New Zealand.

The example is from a new Anthropic report, and it is fun until you imagine a person in place of the pub.

The facts: on 10 September 2026 Anthropic's Frontier Red Team published new evaluations of model capabilities in tactical intelligence targeting and conventional weapons development. Geolocating photos without metadata, reverse image search or tools, across 6,000 photos, Mythos Preview and Mythos 5 reach a median error of 37.0 and 47.2 km and place 23.7% and 23.1% of the photos within 1 km. There is no human baseline on the same photos; as a proxy the company uses data from GeoGuessr players, who look at Street View and can move within the scene: a median error of 151 km for the Champion Division, the top 0.01% of players. In simulated tasks the models write and improve software to guide a quadcopter to a target, to drop a payload and to fly through jammed airspace. In Anthropic's words, for some tasks in the intelligence and military domains models can do things that historically only a small set of highly trained experts could do. The open-weights models that were tested are behind the frontier, typically between Sonnet and Mythos-class models, but often show concerning levels of capability. Anthropic has introduced new classifiers to block such misuse.

The price of anonymity

The key is one sentence in the middle of the text: much of what has protected people from being tracked was not secrecy but cost. The data has been online for a long time. What was expensive was someone sitting down to connect it.

If a model does in minutes work that takes an analyst hours, the cost falls. According to Anthropic, the median sample in one of the evaluations is about 37,000 words, which a person would need about two and a half hours to read. Mythos Preview produces its full assessment in about 11 minutes on average.

Secrecy never protected people as much as expensive labour did.

The text evaluation comes even closer to each of us. From posts online, the models place some people's homes to within a kilometre. In 70% of the successful cases people had given away their location themselves: a dorm, a venue, a street. In 13% the way they write was enough, with the dialect, local media, public transport and sports teams.

Don't post photos from home with a view of the street, and don't put your neighbourhood in your profile. It isn't paranoia. Models already read the details you don't notice, and by Anthropic's count the strongest of them already beat even the best of the people who compete at it.

The visual is generated code art. No third-party images.
Follow usFacebookLinkedIn
Official primary sources
→Anthropic - Measuring tactical intelligence targeting and conventional weapons capabilities of AI models, 10.09.2026
Original: https://wearecoded.com/en/articles/anthropic-geolokaciya-po-snimka-i-dronove.html
ShareFacebookXLinkedInTelegramWhatsApp
← Back to all news