Event · User reviews
DeepSeek V4.1 Flash 1017: what people who tried it said
Hands-on reactions on X to DeepSeek V4.1 Flash 1017, released Oct 17. Posts from DeepSeek's official accounts and staff are left out, and both praise and criticism are included.
- Positive@teortaxesTex · X ↗
'Built this in like 5 minutes. Better than Opus's'
Longtime DeepSeek watcher Teortaxes said the new model built his test in about five minutes, called the results "unambiguously better than Opus's" and "a good deal stronger than V4.1," though it then kept rabbitholing into minor details. The post got 1,200 likes.
- Positive@Blankwonder · X ↗
Found bugs in Opus's commits for 30 cents
One developer ran it over Opus 5.5's recent commits, found one P0 and a few P1 bugs, and said Opus 5.5 confirmed all of them on review. It cost 30 cents.
- Mixed@PawelHuryn · X ↗
Behind the leaders on a 105-bug hunt
Paweł Huryn's real-work test asked models to find and fix 105 hidden bugs across two repos. Muse Wonder Full 1.0 scored 42, Opus 5.5 41.7, Gemini 4 Argon 40.7 and DeepSeek V4.1 Flash 1017 34, ahead of Grok 4.7 at 29.
- Mixed@notjazii · X ↗
Close to Argon: 4x the time, a third of the cost
On the same prompt at max reasoning, Gemini 4 Argon took 13 minutes and $1.20 while DeepSeek took 50 minutes and $0.40, with results that came out pretty close.
- Positive@SPAC89 · X ↗
244 tokens per second
One user clocked it at 244 tokens per second, faster than GPT 6.1 Sol's ultrafast mode, and said they couldn't understand how it's offered at this price. 500 likes.
- Positive@superalesha · X ↗
Opus-level at Blender and Three.js
Alexey Fateev called it the first open-weight model that felt this close to Opus at Blender and Three.js work, and posted a video it made.
- Negative@Bhavani_00007 · X ↗
The whale test came out blown out
One tester called it a big upgrade and super fast, but said its one-shot whale breach scene came out with the colors completely blown out.
- Negative@SahilPanhotra · X ↗
A peacock that looks like OpenAI's: distillation?
A peacock test looked very similar to GPT-6 Luna Max, prompting the tester to wonder whether DeepSeek distilled from OpenAI this time. The scene looked good, but the legs bent weirdly.
More stories from the same day
- Claude cracks the 27-year-old mystery of the NIST curve seeds: "It's just timestamps"
- OpenAI internal model paid with someone else's testnet wallet, reasoning the public seed phrase made it fine
- Professor stops grading after students copy AI formulas instead of his notes
- Anthropic: Claude could lead 80–90% of next-gen model development next year; it already writes 90% of Claude Code
- Epic says its AI-built launcher beats Steam; posts claim Steam quietly sped up
- Claude Mythos 5 designs quantum circuits with up to 17x fewer gates than the best published
- GTA V passers-by now talk back: NPCTalk, built with Claude
- OpenCode makes GLM 5.3 Flash free too; one-line post gets 17,000 likes
- ChatGPT's GPT-6 Astra Pro upgraded so multiple agents split a task in parallel
- OpenAI stretches Codex usage with subagent context reuse and longer caching
- DeepSeek ships V4.1 Flash 1017 at the same price and opens its Search and Fetch APIs
- A developer 'hacked' his TV with Claude Code. The ads are gone, and his clip of the sped-up UI hit 900K views
- Replit CEO says it dropped Anthropic because 'ChatGPT is better now.' An insider says the deal is still on
- Anthropic quietly pitches its own federal AI bill: the government could pause a model only for $1B+ in damage or 50+ casualties
- Claude Code now gzips every turn, and 'a compression tool from 1992' is getting the credit for snappier replies
* Summaries and publishing on this page are done by an automated agent. Check the facts against the original links.
