DeepSeek pushed a point update on October 17, 2026 09. Same price. Same API name. DeepSeek's change log says the new checkpoint, V4.1-Flash-1017, scores 42.3 01 on Terminal-Bench 4.0.
DeepSeek puts the prior checkpoint at 31.2 02. That is 11.1 25 points in one release. Divide 11.1 by 31.2 and you get a 36% 26 jump. DeepSeek also reports its Humanity's Last Exam score without tools rising from 36.8 03 to 42.6.
The rivals are close behind. Artificial Analysis scores DeepSeek V4.1 Flash 1017 (Max) at 45 04 on its Intelligence Index. It scores MiniMax M3.1 Flash and GLM-5.3-Flash at 42 06 each. Three cheap models now sit inside a 3-point band.
Inception claims its diffusion model Mercury 3 runs 4x 13 to 8x faster than GPT-6 Luna and Claude Haiku 5.5, and beats both on benchmarks. When the smart scores bunch up, the contest moves to dollars and seconds.
Most of the agent scores here come from the labs themselves. kie.ai reports the first independent tests of 1017 arrived on October 18. Those tests found GPT-6 Astra still leads on repeated-run harness scores. Treat every vendor number below as a claim you test.
Cost per finished task picks the flash tier
A flash model is the cheap, fast tier a lab sells under its flagship. Once three of them score within 3 points of each other, the number that matters is cost per finished task: the full bill for a batch of runs divided by the runs that actually worked.
What a finished agent task costs across the flash tier
Think of it like paying a courier. You pay for parcels that arrive on time. The miles on the van are the courier's problem, right up until they show up on your invoice.
VentureBeat ran this math in a business-task comparison. It put DeepSeek V4.1-Flash-1017 at about $0.26 28 per completed task and GLM-5.3 Flash at $0.28. Gemini 3.8 Flash came in at $1.24. Divide $1.24 by $0.26 and Gemini costs about 4.8x more per finished task.
Every cost in a flash-tier agent lands in one of four buckets:
- Price per token. DeepSeek says pricing did not change with 1017. kie.ai reports off-peak cached input at $0.003 per million tokens.
- Tokens per task. DeepSeek calls 1017 a performance-cost Pareto improvement, which means better scores without a bigger bill. PC Watch reports it cut both task time and output length.
- Retries. A run that fails and reruns pays twice.
- Seconds. Artificial Analysis measured DeepSeek's flash model at 1.09 seconds to first token.
Retries hide the most money. Chain 100 29 steps at 98% success each, and 0.98 to the 100th power gives about 13% clean runs. Push each step to 99.5% 30 and clean runs climb to about 61%. So a 1.5-point 31 gain per step can matter more than a 3-point index gap.
Inside the 3-point flash spread
The tight index score hides real differences by job. DeepSeek's comparison table shows 77.0 20 on DeepSWE v1.1, a code-repair test, against 74.0 21 for Claude Opus 5. PC Watch reports the same table puts 1017 ahead of Opus 5 on Automation-Bench, 60.5 08 to 50.3 22. kie.ai notes the model still trails Opus 5 on Terminal-Bench 4.0.
MiniMax takes other rounds. Artificial Analysis's comparison puts MiniMax M3.1 Flash at 48% 24 on Humanity's Last Exam. DeepSeek sits at about 43% on that test, with GLM-5.3 Flash at 40%. MiniMax reports 62.2 on Automation-Bench, a hair above DeepSeek's 60.5 08. VentureBeat also flagged a "high hallucination rate" for the MiniMax model.
Picking from a leaderboard is like hiring from a CV. The trial shift tells you more, and your trial shift is the harness.
A harness is the code wrapped around the model, from the system prompt to the sandbox where commands run. Every benchmark above ran inside some harness. Yours is different.
Version drift lives in the harness too. DeepSeek's change log says the deepseek-flash name now points at 1017 automatically. Opper offers a key pinned to the 1017 checkpoint, and its undated key still serves the September weights. If your agent calls the moving name, the model under your tools changed on October 17 without you touching a line of code.
Here is how the money works out. Say your agent finishes 1,000 18 tasks a day. At VentureBeat's per-task figures, DeepSeek costs $260 32 a day and GLM costs $280 33. Gemini 3.8 Flash costs $1,240 34.
Over 30 days 35, DeepSeek runs $7,800 37 and GLM runs $8,400. That $600 36 gap is real, but it is thin. It disappears if DeepSeek needs one extra run on about 8 tasks in every 100, since eight percent of $7,800 is $624.
Speed has a price too. At Inception's claimed 4x 13 to 8x speedup, a 40-second 38 agent step would take 5 to 10 seconds. For a support bot, that is the difference between a customer waiting and a customer leaving. The claim is Inception's own, and we have not seen it reproduced.
We think the right model for your workload is whichever one clears your own test set at the lowest cost per finished task. The index tells you which three to test. Your logs tell you which one to ship.
Where cheap agent budgets quietly leak
Small step failures compound fast.
Chain 100 steps at 98% success each and only about 13% of runs finish clean. Lift each step to 99.5% 30 and clean runs climb to about 61%, so a per-step gain can outweigh a 3-point index gap.
The moving model name changed weights overnight.
DeepSeek's change log says the deepseek-flash name now points at 1017 automatically. Agents calling that name got a new model under their tools on October 17 with no code change.
A new checkpoint every five weeks.
PC Watch reports 1017 arrived 37 days 19 after the September V4.1 Flash. At that pace a team tracking DeepSeek, MiniMax and GLM would face about 30 41 swaps a year, though the next 11 points may come harder.
Harnesses outlast ten checkpoints by 2027
On May 15, 2026, FriendliAI described DeepSeek-V4-Flash as a 284-billion-parameter mixture-of-experts model with 13 billion active. Mixture-of-experts means only a slice of the model runs for each token. PC Watch reports the 1017 checkpoint at 552 billion parameters, with 8 billion active to read input and 16 billion to write output. PC Watch also reports 1017 arrived 37 days 19 after the September V4.1 Flash.
Run that cadence forward. One checkpoint every 37 days 19 works out to about 10 39 a year from one lab, or about 49 40 over five years. If MiniMax and Z.ai's GLM kept a similar pace, a team tracking all three would face about 30 41 swaps a year.
At that rate, the model pick is an asset that depreciates in about five weeks. A test set built from your own tasks does the opposite. It compounds, because every new checkpoint gets scored in an afternoon. Building it costs one weekend. Skipping it costs a blind swap every 37 days 19.
PC Watch reports DeepSeek posted the 1017 weights on Hugging Face under the MIT license the same day. Open weights let other hosts serve the same model, and more hosts means more spread in price and latency behind one model name.
The labs already compete on extras around the model. DeepSeek opened its Search and Fetch APIs to all users on October 17, and PC Watch reports a price of $0.75 18 per 1,000 calls. Inception pitches raw speed. We expect the next year of flash releases to fight over speed and tool add-ons.
Whether the 37-day 19 cadence holds is unclear. Big jumps off a low base come easier than big jumps near the top. Terminal-Bench at 42.3 01 still leaves plenty of headroom, but the next 11 points may cost DeepSeek far more than the last 11 did.
Run a 30-task flash bake-off Saturday
Step one: pull 30 41 real tasks from last month's logs. Pick ones with a clear pass or fail, like a test suite going green or a form filed correctly. Thirty is small enough to run three models in one afternoon.
Step two: pin every model version. Pinning means calling an exact checkpoint so the provider cannot swap it mid-test. For DeepSeek, use a dated key such as the 1017 key Opper offers.
Step three: run DeepSeek V4.1-Flash-1017, MiniMax M3.1 Flash and GLM-5.3-Flash through the same harness. Keep everything else identical, down to the prompt and the tool list. If the work is code, OpenCode 2.0.30 docks terminal diffs next to the session. You can read exactly what each model changed.
Step four: log five fields per run. Record pass or fail, total tokens, tool calls, retries and wall-clock seconds. A spreadsheet is enough.
Step five: divide total spend by passed tasks to get cost per finished task. Then find p95 latency, the time that 95 of every 100 30 runs beat. That number is what your slowest users feel.
Step six: break things on purpose. Kill a tool halfway through a run and feed back one malformed JSON reply. JSON is the structured text format tools pass back to the model. Watch which model recovers and which one keeps going as if nothing happened.
Expect at least one model to fail in a way no benchmark showed. That failure is the most useful row in your sheet. It tells you where to add a check before anything reaches production.
Step seven: save the 30 41 tasks and the script. DeepSeek shipped 1017 37 days 19 after the last checkpoint. Put a rerun on the calendar for November 23. When the next checkpoint lands, you score it in an hour and switch only if the cost per finished task drops.
Run a 30-task flash bake-off on your own logs
- Pull 30 41 real tasks. Take them from last month's logs and keep only ones with a clear pass or fail, like a green test suite or a correctly filed form.
- Pin and run three models. Call dated checkpoints such as the 1017 key Opper offers, then run DeepSeek V4.1-Flash-1017, MiniMax M3.1 Flash and GLM-5.3-Flash through the same prompt and tool list, logging pass or fail, tokens, tool calls, retries and seconds.
- Score cost per finished task and break things. Divide total spend by passed tasks, find p95 latency, then kill a tool mid-run and feed back one malformed JSON reply to see which model recovers. Book a rerun for November 23.
Pin the checkpoint and keep your own scorecard.
Three flash models now land within 3 points on the index, and the gap between DeepSeek and GLM is about $600 36 a month at 1,000 18 tasks a day. A few extra retries erase that gap, and a silent version swap can change everything under your tools. A 30-task 41 test set built from your logs scores each new checkpoint in an afternoon. We would ship whichever model clears it at the lowest cost per finished task, then rerun it every time a lab pushes an update.