K Koda Intelligence
DEEP DIVE · STRATEGY

Food Delivery Firm Trails Claude by 0.7

Meituan's LongCat-2.5 scores 88.4 on Terminal-Bench v2.1, close behind Claude Opus 5 at 89.1, by Meituan's own figures. The weights are free under an MIT license. Serving 1.6 trillion parameters took 194 Huawei Ascend chips in Meituan's sample deployment. On the harder v4.0 test, Opus 5 still leads by 15.4 points. We think the durable edge now sits in your workflow, your data and your integrations.

By the Koda editorial team · · 6 min read

Meituan sells food delivery. Its LongCat team just released an agent model that scores 88.4 on Terminal-Bench v2.1,1 according to Meituan's own figures. Claude Opus 5 scores 89.1 on the same table2. GPT-5.6 Sol scores 88.8.

That leaves LongCat-2.5 0.7 points behind Anthropic's Claude Opus 5 and 0.4 points behind OpenAI's3. MLQ.ai reports that Meituan put the weights on Hugging Face on November 9, 2026,4 about a week after unveiling the model. The license is MIT. Anyone can modify the model and sell products built on it.

Most teams still cannot run it. LongCat-2.5 has 1.6 trillion1 parameters. The Decoder reports it was trained entirely on Chinese hardware, and that Meituan's sample deployment uses 194 Huawei Ascend chips2. The download is free, and the hardware bill is huge.

A delivery app's lab can now land within one point of the frontier. If that is true, the model you rent is a weak moat. Your workflow, your data and your integrations have to carry the weight.

Easy tests converge, hard tests split

Meituan's numbers tell two stories. Terminal-Bench v2.1 tests agents on command-line and coding work. On it, the four models in Meituan's table sit within 2.2 points, with DeepSeek-V4.1-Flash on top at 90.6.5

What a free download still costs to run and rentVENTUREBEAT · THE DECODER · MEITUAN
LongCat-2.0 input price1VentureBeat · per million input tokens, standard API$0.75
LongCat-2.0 output price1VentureBeat · per million output tokens, standard API$2.95
Accelerators to serve it2The Decoder · Huawei Ascend chips, sample deployment194
Four-bit prefill throughput gainMeituan · versus an eight-bit baseline43%

Terminal-Bench v4.0 is harder, and the pack breaks apart. Meituan's figures put Claude Opus 5 at 53.9.2 LongCat-2.5 scores 38.5,6 edging past GPT-5.6 Sol's 37.3 by 1.2 points7. The gap to Opus grows from 0.7 points to 15.4.3

Once a test is known, it turns into a commodity within months. The edge then moves to the tasks the test cannot yet measure. The Decoder says the same thing more plainly: LongCat-2.5 performs at the level Claude Opus 5 reached "a few months ago" on coding benchmarks.

The day's facts fall into two piles. The first pile is cheap and getting cheaper. It holds the weights under an MIT license and a v2.1 score that several labs now match.

The second pile is scarce. Opus 5 still leads the hardest tasks by 15.4 points. Serving 1.6 trillion1 parameters on 194 accelerators takes real engineering2.

The scarcest items never ship with any model download: your failure logs and the systems your agent is allowed to touch.

Meituan's real asset is the delivery network

A frontier lab earns its money selling model access, so its margin lives in the model. Meituan earns its money moving orders from merchant to doorstep. To Meituan a model is a cost center. It can give one away and lose nothing it sells.

That is counterpositioning, and Meituan has run the play before. VentureBeat reported on June 29, 2026 that LongCat-2.0 had spent two months as "Owl Alpha," an anonymous model leading developer charts on OpenRouter8. Its standard API price was $0.75 per million input tokens and $2.95 per million output tokens1, with cache hits free. The weights back then were still "coming soon."

LongCat-2.5 closed the loop about four months later, when the weights shipped under MIT. Competitors can now download Meituan's model. They cannot download Meituan's records of how food moves through Chinese cities.

The engineering notes point the same way. Meituan reports that quantizing expert weights and activations to four bits lifts prefill throughput about 43% over an eight-bit baseline. Prefill is the step where the model reads the prompt before it writes a word. Meituan put its effort into serving the model cheaply on Chinese chips. That is where its own costs sit.

Today the evidence supports one decision: treat the model as a part you will swap, and build the swap in from day one. We think teams that hard-wire one vendor's prompt formats and tool schemas are making a bet the last five months argued against.

Two findings would change that call. First, independent testing could show the v2.1 scores do not hold. These are Meituan's figures, and Terminal-Bench's own published work runs several agent scaffolds, a reminder that the wrapper moves the score. Second, the 15.4-point v4.0 gap could survive the next two releases, which would keep model choice decisive for hard work.

Security adds a cost the benchmark hides. Australia's cyber-security guidance warns that open weights add duties for monitoring and supply-chain security. It notes that tampered weights can carry backdoors that ordinary testing may miss. Give an agent write access to production, and that stops being a research worry. It becomes an operations risk.

It is unclear whether a small team can carry those duties for a model this size. For most builders, the realistic path is a hosted version of an open model with their own tests wrapped around it.

Model choice may matter less by 2031

The trend line is short but steep. VentureBeat dated LongCat-2.0's unveiling to June 29, 2026.8 MLQ.ai dated the LongCat-2.5 weights to November 9, 2026.4 That is about 19 weeks between major releases from one lab.

Five years is about 260 weeks. At one release every 19 weeks, Meituan alone would ship about 13 more major models by late 2031. DeepSeek already tops Meituan's own v2.1 table, so the pace across open labs is faster still.

On v2.1 the gap is already under one point. If the lag of "a few months" that The Decoder describes holds, today's hard test becomes next year's commodity test. A harder test takes its place, and the race restarts at the top.

The closed labs may end up holding a lead only on the hardest, newest tasks. How long that lead lasts is an open question. A 15.4-point gap on v4.0 is large today, and we have only one release to measure it against.

The risk is lopsided. A team that builds model swapping into its stack pays a few weeks of engineering. If closed labs keep their edge, that team keeps using them for hard tasks and loses little. If open models keep catching up, a team locked to one vendor loses pricing leverage on every token it buys.

What compounds sits outside the model. Every logged failure sharpens your test set. Every integration with a payment or inventory system raises the cost of leaving. That flywheel belongs to you, whatever model sits in the middle.

Write a private agent test before switching

Step one: pull 25 real tasks from your own logs. Pick work your agent already does, like closing a support ticket or running a database migration. Write down what "done" looks like for each one in a single checkable sentence.

Step two: turn those tasks into an eval. An eval is a fixed test set you rerun every time you change a model or a prompt. Store each task with its input and a short script that checks the result.

Step three: run all 25 tasks through two models. Use your current model and one open-weight model through a hosted API, since few teams can stand up 194 accelerators2. Run each task three times, because agents fail unevenly.

Step four: compute cost per successful task by adding the token cost of every run and dividing by the runs that passed. At VentureBeat's standard LongCat-2.0 rates, a task that reads 200,000 tokens and writes 20,000 costs $0.15 plus $0.059, about $0.21. If it passes 8 of 10 runs, each success costs about $0.26. Prices for LongCat-2.5 may differ, so plug in whatever your provider charges.

Step five: price the failures. A failed run that leaves a ticket open costs a few minutes of human time. A failed run that issues a wrong refund costs real money. Weight each failure by what it would cost you, and compare the two models on that number.

Expect the first pass to break. Tool-call formats differ between models, so early failures are often plumbing. Fix the adapter and rerun the set.

Keep every failure in one file. That file is the start of the proprietary data this whole argument rests on. Rerun the set every Friday. When the next open release lands, you will know within an hour whether it earns a slot in your stack.

Practise it step by step: Design an AI-Assisted Workflow

The bottom line

Rent the model and own the flywheel.

Meituan can give its model away because its money comes from moving food across cities. Builders can borrow that logic. Treat whichever model you use as a replaceable part, and invest in the failure logs, evals and integrations no download can copy. A few weeks of swap engineering is cheap insurance whichever way the frontier gap moves.

LISTEN · AUDIO BRIEFINGThe conversation · ~11 min
WATCH · VISUAL NARRATIVEAnimated breakdown · ~7 min
PLAY · YOUTUBE

Sources

Numbers in the text point here. Verified means we confirmed it against the source; Reported means the source states it and we could not confirm it independently.

  1. 1longcat.ai · LongCat-2.5 scores 88.4 on Terminal-Bench v2.1, according to Meituan's own figures.4 verified, 2 reported
  2. 2the-decoder.com · Claude Opus 5 scores 89.1 on Terminal-Bench v2.1, according to Meituan's benchmark table.3 verified, 2 reported
  3. 3anthropic.com · That leaves LongCat-2.5 0.7 points behind Anthropic's Claude Opus 5 and 0.4 points behind OpenAI's.Reported
  4. 4mlq.ai · MLQ.ai reports that Meituan put the LongCat-2.5 weights on Hugging Face on November 9, 2026.Verified
  5. 5api-docs.deepseek.com · DeepSeek-V4.1-Flash tops Meituan's Terminal-Bench v2.1 table with a score of 90.6.Reported
  6. 6huggingface.co · LongCat-2.5 scores 38.5 on Terminal-Bench v4.0, according to Meituan's figures.Verified
  7. 7benchgap.net · GPT-5.6 Sol scores 37.3 on Terminal-Bench v4.0, according to Meituan's figures.Verified
  8. 8venturebeat.com · VentureBeat reported on June 29, 2026 that Meituan's LongCat-2.0 had spent two months as "Owl Alpha," an anonymous...2 verified, 1 reported

How this piece was checked

Before publishing we pulled 34 claims out of the draft and checked each one against live sources: 14 verified, 10 reported by a named source and 10 our own calculations.

See all 34 claims
  1. 01LongCat-2.5 scores 88.4 on Terminal-Bench v2.1, according to Meituan's own figures.longcat.aiVerified
  2. 02Claude Opus 5 scores 89.1 on Terminal-Bench v2.1, according to Meituan's benchmark table.the-decoder.comReported
  3. 03GPT-5.6 Sol scores 88.8 on Terminal-Bench v2.1, according to Meituan's benchmark table.the-decoder.comVerified
  4. 04Meituan's LongCat team released an agent model called LongCat-2.5.longcat.aiVerified
  5. 05That leaves LongCat-2.5 0.7 points behind Anthropic's Claude Opus 5 and 0.4 points behind OpenAI's.anthropic.comReported
  6. 06MLQ.ai reports that Meituan put the LongCat-2.5 weights on Hugging Face on November 9, 2026.mlq.aiVerified
  7. 07The Decoder reports that LongCat-2.5 was trained entirely on Chinese hardware.the-decoder.comReported
  8. 08VentureBeat reported on June 29, 2026 that Meituan's LongCat-2.0 had spent two months as "Owl Alpha," an anonymous model.venturebeat.comReported
  9. 09As of VentureBeat's June 29, 2026 report, the LongCat-2.0 weights were still listed as "coming soon."venturebeat.comVerified
  10. 10VentureBeat dated the unveiling of Meituan's LongCat-2.0 to June 29, 2026.venturebeat.comVerified
  11. 11Meituan is a food delivery company.meituan.comReported
  12. 12Meituan released the LongCat-2.5 weights about a week after unveiling the model.huggingface.coVerified
  13. 13The LongCat-2.5 weights shipped under the MIT license about four months after LongCat-2.0's June 29, 2026 unveiling.longcat.aiReported
  14. 14LongCat-2.5 has 1.6 trillion parameters.longcat.aiReported
  15. 15According to The Decoder, Meituan's sample deployment of LongCat-2.5 uses 194 Huawei Ascend chips.the-decoder.comVerified
  16. 16Australia's cyber-security guidance warns that open-weight AI models add duties for monitoring and supply-chain security.cyber.gov.auReported
  17. 17Australia's cyber-security guidance notes that tampered open model weights can carry backdoors that ordinary testing may miss.cyber.gov.auVerified
  18. 18DeepSeek-V4.1-Flash tops Meituan's Terminal-Bench v2.1 table with a score of 90.6.api-docs.deepseek.comReported
  19. 19Claude Opus 5 scores 53.9 on Terminal-Bench v4.0, according to Meituan's figures.the-decoder.comVerified
  20. 20LongCat-2.5 scores 38.5 on Terminal-Bench v4.0, according to Meituan's figures.huggingface.coVerified
  21. 21GPT-5.6 Sol scores 37.3 on Terminal-Bench v4.0, according to Meituan's figures.benchgap.netVerified
  22. 22According to VentureBeat, LongCat-2.0's standard API price was $0.75 per million input tokens.longcat.aiVerified
  23. 23According to VentureBeat, LongCat-2.0's standard API price was $2.95 per million output tokens.longcat.aiVerified
  24. 24According to VentureBeat, LongCat-2.0 API cache hits were free.Reported
  25. 25LongCat-2.5 is 0.7 points behind Claude Opus 5 on Terminal-Bench v2.1 (89.1 minus 88.4).Koda calculationOur math
  26. 26LongCat-2.5 is 0.4 points behind OpenAI's GPT-5.6 Sol on Terminal-Bench v2.1 (88.8 minus 88.4).Koda calculationOur math
  27. 27The four models in Meituan's Terminal-Bench v2.1 table sit within 2.2 points of each other (90.6 minus 88.4).Koda calculationOur math
  28. 28LongCat-2.5 leads GPT-5.6 Sol by 1.2 points on Terminal-Bench v4.0 (38.5 minus 37.3).Koda calculationOur math
  29. 29LongCat-2.5's gap to Claude Opus 5 grows from 0.7 points on Terminal-Bench v2.1 to 15.4 points on Terminal-Bench v4.0 (53.9 minus 38.5).Koda calculationOur math
  30. 30About 19 weeks passed between Meituan's LongCat-2.0 unveiling (June 29, 2026) and the LongCat-2.5 weights release (November 9, 2026).Koda calculationOur math
  31. 31Five years is about 260 weeks.Koda calculationOur math
  32. 32At one release every 19 weeks, Meituan would ship about 13 more major models by late 2031.Koda calculationOur math
  33. 33At LongCat-2.0's standard rates, a task that reads 200,000 tokens and writes 20,000 tokens costs $0.15 plus $0.059, about $0.21.Koda calculationOur math
  34. 34A $0.21 task that passes 8 of 10 runs costs about $0.26 per successful run.Koda calculationOur math

How we check · Audit KODA-20261116-8F953B9FC2EB

Browse the Deep Dive archive

Get the morning Signal

233 editions so far, one a day. Unsubscribe anytime.