DeepSeek Flash vs Claude Opus 5 at 35.4 and 28.6

DeepSeek released an open-weights Flash model with a side-by-side Agents' Last Exam line of 35.4 for that Flash checkpoint and 28.6 for Claude Opus 5, on DeepSeek's own tables. Opus 5 is the closed model Anthropic sells into that comparison.

Posts from @k2sbhai and @imjustnewatai circulating the table treat 35.4 as beating Claude Opus 5 at coding. The axis they skip is who wrote the scoreboard: 35.4 and 28.6 are one DeepSeek-published Agents' Last Exam pair, and Claude Opus 5 is still Anthropic's model.

35.4 against 28.6 on Agents' Last Exam

Subtract the published pair and the gap is 6.8 points: 35.4 for the open-weights Flash model minus 28.6 for Claude Opus 5, both taken from DeepSeek's tables.

35.4 for the open-weights Flash model and 28.6 for Claude Opus 5, on Agents' Last Exam, as printed by DeepSeek.

6.8 points is the gap on the Agents' Last Exam pair DeepSeek printed. Posts from @k2sbhai and @imjustnewatai frame that table as the Flash model beating Claude Opus 5 at coding. DeepSeek printed the pair for the open-weights Flash model at 35.4 and for Claude Opus 5 at 28.6. Your project is one task family, and families differ. Ticket triage, repository edits, and a stop condition are a coding-agent loop. Drafting, meeting notes, and a product UI are a different loop. The Agents' Last Exam published pair is a direct measurement for the coding-agent loop only to the extent the items resemble that loop, and a neighboring measurement for drafting and UI work. Opus 5 can be the production default on the neighboring work while the DeepSeek table shows 28.6. Flash can be the checkpoint you pin for repository edits while your own logs are still the record that counts. Use the row as a prior. Update it when those logs disagree.

35.4 and 28.6 are one row from one lab.

$0.15 to $0.25 per task

DeepSeek's open-weights Flash release is dated October 17, 2026, and the Agents' Last Exam head-to-head with Claude Opus 5 is the published pair on that drop. The 35.4 and 28.6 figures name both models on one DeepSeek table. Cost sits on a separate line from that exam.

Posts from @k2sbhai and @imjustnewatai fold the published pair into one line about beating Opus at coding. Agents' Last Exam, printed beside Opus 5, is the head-to-head DeepSeek put in front of readers. A model can lead on its publisher's agent exam and still be an awkward default inside a product that only calls Claude. A closed flagship can trail on one published agent exam and still be the model a company already cleared for customer data. The two outcomes can be true on the same afternoon, because they answer different constraints.

The open-weights Flash model is the open-weights side of the pair. The checkpoint is published as open weights. You can pin a revision, place it on hardware you control, and keep that file when a hosted API later changes its alias. Opus 5 is served as a hosted Anthropic model. One promise is a file you can hash. The other is an endpoint someone else runs. The exam points and the operational promise answer different questions, and a buyer who mixes them will brief the wrong meeting.

Hold the published pair in view when you brief a team. 35.4 says how the open-weights Flash model scored on Agents' Last Exam beside the 28.6 DeepSeek listed for Opus 5. A coding agent that spends an hour in a repository produces a long trace. An exam prints the published pair. Teams that already track success on their own repos should keep that log, and treat DeepSeek's table as the public prior for this checkpoint. The prior is specific to the exam named on the card. It is the pair DeepSeek published for this Flash build against this Opus model.

The table is what fits on a phone. 35.4 against 28.6 is a subtraction anyone can do, beside the open-weights Flash checkpoint. Compression is why the pair travels as a coding verdict. Sorting it back out is the work. One column is a file you can hash, released as open weights. One column is Anthropic's hosted Opus 5. The grader for both numbers in the head-to-head is DeepSeek. Once those three facts are in order, the screenshot is a useful summary of a single lab's row.

Brief the October 17, 2026 row as three separate lines. Line one is 35.4 for the open-weights Flash model. Line two is 28.6 for Claude Opus 5. Line three is the 6.8 point gap between them, graded by DeepSeek. Then brief the deliverable. Open weights mean a file the buyer can place on hardware they control. Hosted Opus 5 means an Anthropic endpoint. @k2sbhai and @imjustnewatai put the first three lines in front of people as a coding comparison. The fourth line, file or endpoint, is what changes the purchase order. Leave the exam cells on the slide. Put the deliverable in the recommendation.

Keep the bill in its own column. The Agents' Last Exam pair is a quality comparison DeepSeek published. Cost on this matchup is $0.15 to $0.25 per task for the open-weights Flash model and $10 or more per task for Opus. A fair cost pass multiplies your trace length by the rate card you actually hold and adds operator time for the open checkpoint. Long traces and short chat turns can favor different sides of that worksheet. Bring the trace length from your logs. Leave 35.4 and 28.6 as exam cells.

2 lanes for a file and a hosted model

Choose the open-weights Flash model when the thing you need is a file. Open weights are that file. The exam context DeepSeek attached is 35.4 against Claude Opus 5 at 28.6. The file lane fits a developer or a technical founder who will run inference, accept the server, and wants the weights even when a hosted Claude workspace would take fewer clicks. It also fits a vendor who must hand a customer a model file inside a VPC, where an outside API is the thing the contract forbids.

Choose Claude Opus 5 when the thing you need is Anthropic's hosted model, with Anthropic's changelog and Anthropic's API. On DeepSeek's Agents' Last Exam table the hosted choice is the 28.6. The hosted lane fits a company already standardized on Claude, a builder shipping inside a product that calls Anthropic, and anyone whose paperwork names Anthropic and will reject a self-hosted weight. For a coding-agent product, this is the setup constraint. Flash gives you a checkpoint to mount behind your own router. Opus 5 gives you the endpoint customers already approved. The October 17, 2026 table has Flash ahead by 6.8 on the exam DeepSeek graded. The hosting rule decides which of those models you are allowed to propose.

Use 35.4 against 28.6 when the question is this specific Opus 5 pairing. For a non-engineer, the practical fork is the account you can open this week. A builder product that asks for a Claude key is already pointed at Opus 5's lane. A builder product that asks for a base URL or a local model is pointed at the open-weights Flash file. The scores say the open checkpoint belongs in the same conversation as Opus 5. Someone still has to configure inference, or the file stays a download.

A non-coder building a project still has to pick a lane before the 35.4 matters. If the tool in front of them asks for a Claude account, the path is Opus 5 and the DeepSeek cell beside it is 28.6. If the tool asks them to point at a model file, the path is the open-weights Flash checkpoint and the cell beside it is 35.4. The 6.8 point gap tells them the open checkpoint is ahead on the exam DeepSeek printed. The account or the file is the step that makes either score usable.

A downloadable Flash revision and a hosted Opus 5 are different procurement objects, so the open checkpoint changes the buy on the terms of the file itself.

October 17, 2026 is the date on DeepSeek's Agents' Last Exam pair, with the open-weights Flash model at 35.4 and Claude Opus 5 at 28.6. The gap on that pair is 6.8 points.