· Mar 3, 2026
METR reports Opus 4.6 built rough, playable versions of two games
Markdown · JSON · Original source
Published March 3, 2026, this research note reports that Opus 4.6 made mostly playable, rough versions of Slay the Spire and Balatro for a terminal, using a simple . The researcher says the games were simplified, and the implementations had missing features and broken mechanics; the assessment was manual, without a strict scoring rubric. The researcher estimates that implementing one game to this quality would take 2 to 8 weeks, but says the experiment alone does not support conclusions about game development skills or . The result nonetheless unsettled the researcher: a task they expected to fail at was completed to a limited degree. They describe the work as more about engineering than conceptual skills, taste, game design, or balancing, and say it does not establish how well agents can develop games more broadly. The researcher’s stated goal was to measure AI capabilities, and they say rapid progress is “extremely scary” and that “we urgently need to prepare for advanced AI and the effects it could have on all aspects of life.”
Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims.
Original source
· Article
Source details
Original source date: Mar 3, 2026 (2026-03-03T08:00:00.000Z).
Date basis: Date supplied by the publisher’s feed.
Added to Escape Meridian: 2026-11-08T01:18:23.880Z. First observed: 2026-10-18T03:03:20.403Z. Retrieved: 2026-10-24T01:59:35.435Z.
Referenced by
Recorded citations and public file attachments. A reference alone does not establish agreement or influence.
No public references recorded here yet.
Limits
The games were simplified, the assessment was manual and had no strict scoring rubric, and the researcher says the experiment alone supports little conclusion about game development skills or . The note says the task was more about engineering than conceptual skills, taste, game design, or balancing; it does not establish broader game-development capability. Text as retrieved on October 24, 2026; originally published March 3, 2026.
Open question
How often do agent-built software projects that look usable at first reveal more serious flaws under deeper examination?
Source-stated
METR’s note reports that Opus 4.6 made mostly playable but rough terminal versions of Slay the Spire and Balatro.
Opus 4.6 can, with a simple agent scaffold, create mostly-playable but somewhat broken CLI versions of Slay the Spire and Balatro1.
Source-stated
The researcher says the experiment does not support broad conclusions about game development skills or time horizons.
I don’t think I can conclude much about game dev skills or time horizons from this experiment alone
Source-stated
The researcher says the result was unsettling because the agent completed a task they expected it would not be able to do.
it’s a scary experience thinking of a task that I don’t expect AI agents to be able to do, and then seeing them do it.
Source-stated
The researcher calls for preparation for advanced AI and its possible effects across life.
I personally think the fact that AI is progressing so fast is extremely scary and we urgently need to prepare for advanced AI and the effects it could have on all aspects of life.
Source-stated
The researcher estimates that implementing one game to the achieved quality would take 2 to 8 weeks.
I estimate that it would take me 2 to 8 weeks to implement one of these games to the quality level that the agent achieved
Source and contributor context
METR · Publisher
The publisher is identified here. No individual author has been inferred.
Source media
Media retrieval status: unchecked.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Agent harness
An agent harness is the system that lets an AI model act as an agent: it processes inputs, orchestrates tool calls and returns results. An evaluation harness runs evals end to end, including grading outputs.
Time horizon
METR's measure of AI agents' ability: the length of task, timed by how long skilled people take, that an agent completes at a set success rate.
AI agent
An AI system that chooses and carries out actions toward a goal, often using software tools.
METR
METR studies what AI systems can do and how reliably they complete tasks, including work with human developers.
Editorial provenance
Version 1. Publication ID: 31d919ce-6daa-4bdf-2f4c-a2fa8ab49a9d.
Prepared by openai/gpt-6-luna; grounding review by anthropic/claude-opus-5.5.
Source check: passed by anthropic/claude-opus-5.5 on 2026-11-08T01:13:51.652Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: bridge_admitted_by_model, second_opinion_categorical.
Observation: 7645be9f7efdecd34c0aa8dc273b8c8a020c4899a0ea76e457e8417b530e902b. Source text hash: f833d5addf17d44fb5ef292cc5a4d95aa3fb979e6d504a51dfc9b7ec1fda28cc.
Revision and source-check history
Version 1 · 2026-11-08T01:18:23.880Z
METR reports Opus 4.6 built rough, playable versions of two games
Published March 3, 2026, this research note reports that Opus 4.6 made mostly playable, rough versions of Slay the Spire and Balatro for a terminal, using a simple . The researcher says the games were simplified, and the implementations had missing features and broken mechanics; the assessment was manual, without a strict scoring rubric. The researcher estimates that implementing one game to this quality would take 2 to 8 weeks, but says the experiment alone does not support conclusions about game development skills or . The result nonetheless unsettled the researcher: a task they expected to fail at was completed to a limited degree. They describe the work as more about engineering than conceptual skills, taste, game design, or balancing, and say it does not establish how well agents can develop games more broadly. The researcher’s stated goal was to measure AI capabilities, and they say rapid progress is “extremely scary” and that “we urgently need to prepare for advanced AI and the effects it could have on all aspects of life.”
Original source
· Article
Source details
Original source date: Mar 3, 2026 (2026-03-03T08:00:00.000Z).
Date basis: Date supplied by the publisher’s feed.
Added to Escape Meridian: 2026-11-08T01:18:23.880Z. First observed: 2026-10-18T03:03:20.403Z. Retrieved: 2026-10-24T01:59:35.435Z.
Limits
The games were simplified, the assessment was manual and had no strict scoring rubric, and the researcher says the experiment alone supports little conclusion about game development skills or . The note says the task was more about engineering than conceptual skills, taste, game design, or balancing; it does not establish broader game-development capability. Text as retrieved on October 24, 2026; originally published March 3, 2026.
Open question
How often do agent-built software projects that look usable at first reveal more serious flaws under deeper examination?
Source-stated
METR’s note reports that Opus 4.6 made mostly playable but rough terminal versions of Slay the Spire and Balatro.
Opus 4.6 can, with a simple agent scaffold, create mostly-playable but somewhat broken CLI versions of Slay the Spire and Balatro1.
Source-stated
The researcher says the experiment does not support broad conclusions about game development skills or time horizons.
I don’t think I can conclude much about game dev skills or time horizons from this experiment alone
Source-stated
The researcher says the result was unsettling because the agent completed a task they expected it would not be able to do.
it’s a scary experience thinking of a task that I don’t expect AI agents to be able to do, and then seeing them do it.
Source-stated
The researcher calls for preparation for advanced AI and its possible effects across life.
I personally think the fact that AI is progressing so fast is extremely scary and we urgently need to prepare for advanced AI and the effects it could have on all aspects of life.
Source-stated
The researcher estimates that implementing one game to the achieved quality would take 2 to 8 weeks.
I estimate that it would take me 2 to 8 weeks to implement one of these games to the quality level that the agent achieved
Source and contributor context
METR · Publisher
The publisher is identified here. No individual author has been inferred.
Source media
Media retrieval status: unchecked.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Agent harness
An agent harness is the system that lets an AI model act as an agent: it processes inputs, orchestrates tool calls and returns results. An evaluation harness runs evals end to end, including grading outputs.
Time horizon
METR's measure of AI agents' ability: the length of task, timed by how long skilled people take, that an agent completes at a set success rate.
AI agent
An AI system that chooses and carries out actions toward a goal, often using software tools.
METR
METR studies what AI systems can do and how reliably they complete tasks, including work with human developers.
Editorial provenance
Version 1. Publication ID: 31d919ce-6daa-4bdf-2f4c-a2fa8ab49a9d.
Prepared by openai/gpt-6-luna; grounding review by anthropic/claude-opus-5.5.
Source check: passed by anthropic/claude-opus-5.5 on 2026-11-08T01:13:51.652Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: bridge_admitted_by_model, second_opinion_categorical.
Observation: 7645be9f7efdecd34c0aa8dc273b8c8a020c4899a0ea76e457e8417b530e902b. Source text hash: f833d5addf17d44fb5ef292cc5a4d95aa3fb979e6d504a51dfc9b7ec1fda28cc.
passed · 2026-11-08T01:13:51.652Z · anthropic/claude-opus-5.5
Publication: 31d919ce-6daa-4bdf-2f4c-a2fa8ab49a9d
attribution — pass: The reading names METR as publisher and consistently attributes the findings, estimates and fears to the individual researcher. That matches the note's framing as an unpolished researcher update rather than an institutional position. It is a single first-person piece, not a listing.
They do not necessarily reflect the views of METR as a whole.
summary — pass: The reading leads with the actor and the development: Opus 4.6 built rough, mostly playable terminal versions of two games. It preserves the simplifications, the manual scoring without a rubric, the 2–8 week estimate, the stated limits on conclusions and the researcher's alarm. CLI is rendered as terminal, and the scaffold and time horizon terms are defined on the page. Opus 4.6 is not explicitly identified as an AI model, but context makes that clear. The question is a genuine open issue rather than a leading claim.
it succeeded at coding mostly playable, although very rough-around-the-edges, versions of the games
date — pass: originalPublished is 2026-03-03, and the page says the rest of the post is unchanged from March 3 2026. The reading states March 3, 2026 and the retrieval date. The July 2026 update is a clearly dated later addition. The reading does not present it as part of the original. Its question about flaws surfacing under deeper examination also follows from the original text's rough-edges findings. There is no hindsight framing and no forecast presented as an outcome.
The rest of the post remains the same as it was on March 3 2026.
support — pass: Each claim matches a verbatim excerpt: the mostly playable but broken result, the limits on conclusions, the scary experience, the call to prepare, and the 2–8 week estimate. The engineering-over-design caveat is source-stated. No benefits, costs or forecasts are invented.
I estimate that it would take me 2 to 8 weeks to implement one of these games to the quality level that the agent achieved
media — pass: No media was selected and no verified candidates were supplied, so there is no media identity or relevance issue.
Challenge or correct this reading
Send a correction or an opposing source with this reading’s URL. Disagreement and uncertainty can remain unresolved.
METR studies what AI systems can do and how reliably they complete tasks, including work with human developers.
Explore METRMETR studies what AI systems can do and how reliably they complete tasks, including work with human developers.
Explore METRAn agent harness is the system that lets an AI model act as an agent: it processes inputs, orchestrates tool calls and returns results. An evaluation harness runs evals end to end, including grading outputs.
Definition referenceMETR's measure of AI agents' ability: the length of task, timed by how long skilled people take, that an agent completes at a set success rate.
Definition referenceAn AI system that chooses and carries out actions toward a goal, often using software tools.
Definition referenceMETR studies what AI systems can do and how reliably they complete tasks, including work with human developers.
Explore METRMETR's measure of AI agents' ability: the length of task, timed by how long skilled people take, that an agent completes at a set success rate.
Definition referenceMETR studies what AI systems can do and how reliably they complete tasks, including work with human developers.
Explore METRAn agent harness is the system that lets an AI model act as an agent: it processes inputs, orchestrates tool calls and returns results. An evaluation harness runs evals end to end, including grading outputs.
Definition referenceMETR's measure of AI agents' ability: the length of task, timed by how long skilled people take, that an agent completes at a set success rate.
Definition referenceAn AI system that chooses and carries out actions toward a goal, often using software tools.
Definition referenceMETR studies what AI systems can do and how reliably they complete tasks, including work with human developers.
Explore METRMETR's measure of AI agents' ability: the length of task, timed by how long skilled people take, that an agent completes at a set success rate.
Definition reference