Cursor AI releases Composer 2.5, its most advanced coding model that scores 62 on the Artificial Analysis Coding Agent Index at 10-60 times lower cost than leading rivals
The launch begins expanded scaling collaboration with SpaceXAI.
#23Sasha Rush@SRUSH_NLPHuh, so you’re saying it’s smart, cheap, and fast.
#23Sasha Rush@SRUSH_NLPBeen working on text feedback / OPSD in Composer. Really interesting space, and a much more to be explored.
@eliebakouch Unfortunately can’t give a precise answer. Both were scaled significantly.
elie@eliebakouch@srush_nlp really cool work congrats to the team. if you can answer, do you have rough estimate on how the compute was allocated between RL and continual pt in composer 2 -> 2.5?
#23Sasha Rush@SRUSH_NLPBeen working on text feedback / OPSD in Composer. Really interesting space, and much more to be explored.
@adityagrover_ @siyan_zhao Thanks for publishing! We like your paper a lot.
Siyan, I think we had a chance to meet at NeurIPS. Let us know if you ever are in NYC and want to come by and chat.
Aditya Grover@adityagrover_.@siyan_zhao will be talking more about our work introducing OPSD at ICML in July. https://siyan-zhao.github.io/blog/2026/opsd/
#68clem 🤗@CLEMENTDELANGUEVery cool to see Cursor doubling down on training great models. In my opinion, ultimately all serious companies in AI will want to train models themselves, based on open-source instead of outsourcing AI to others via APIs!
Cursor@cursor_aiIntroducing Composer 2.5, our most powerful model yet. It's more intelligent, better at sustained work on long-running tasks, and more reliable at following complex instructions. For the next week, we’re doubling the included usage of the model.
#76Elon Musk@ELONMUSKTry Composer 2.5 on Cursor!
Michael Truell@mntruellComposer 2.5 is now the most-chosen model in Cursor. We're giving everyone 10x usage for the rest of the day. Enjoy!
#76Elon Musk@ELONMUSKTry it out!
(Partially trained on Colossus 2)
Cursor@cursor_aiIntroducing Composer 2.5, our most powerful model yet. It's more intelligent, better at sustained work on long-running tasks, and more reliable at following complex instructions. For the next week, we’re doubling the included usage of the model.
#76Elon Musk@ELONMUSKTry Composer 2.5
#83rohan anil@_AROHAN_Same pretrain base but little more intelligent on the tasks. Very cool work with Muon (yes, I didn’t bring up Shampoo b2=0.0 yet), cool to see textual feedback in RL.
Cursor@cursor_aiIntroducing Composer 2.5, our most powerful model yet. It's more intelligent, better at sustained work on long-running tasks, and more reliable at following complex instructions. For the next week, we’re doubling the included usage of the model.
#167Emad@EMOSTAQUEThis is such an interesting chart layout, like it a lot!
Congrats to @cursor_ai team on the 2.5 launch 🚀
@elonmusk Is it accessible through the SuperGrok subscription too?
Elon Musk@elonmuskTry it out! (Partially trained on Colossus 2)
#357Boris Power@BORISMPOWERGreat work - exciting to see you training a very powerful coding model!
Cursor@cursor_aiIntroducing Composer 2.5, our most powerful model yet. It's more intelligent, better at sustained work on long-running tasks, and more reliable at following complex instructions. For the next week, we’re doubling the included usage of the model.
#413Aditya Grover@ADITYAGROVER_OPSD now being used at frontier scale to train Composer 2.5. Great work by @srush_nlp and team!
My feed has been filled with commentary in the last few days speculating the behavior of OPSD at scale. Refreshing to see some verification🙂
Sasha Rush@srush_nlpBeen working on text feedback / OPSD in Composer. Really interesting space, and much more to be explored.
.@siyan_zhao will be talking more about our work introducing OPSD at ICML in July. https://siyan-zhao.github.io/blog/2026/opsd/
Aditya Grover@adityagrover_OPSD now being used at frontier scale to train Composer 2.5. Great work by @srush_nlp and team! My feed has been filled with commentary in the last few days speculating the behavior of OPSD at scale. Refreshing to see some verification🙂
Cursor’s Composer 2.5 stirred up the coding war.
Now we have 3 labs capable of training strong coding models: Anthropic, OpenAI, SpaceX (+Cursor).
Wouldn’t be surprised if Google drops a strong coding model tomorrow at I/O.
This is the chatbot war all over again: OpenAI leads, then the market gets divided by other AI labs. Same thing is happening to coding models.
#506Michael Truell@MNTRUELLComposer 2.5 is a significant step up from Composer 2.
This is the very start of our work with SpaceXAI. Hope to have more improvements out soon.
Cursor@cursor_aiIntroducing Composer 2.5, our most powerful model yet. It's more intelligent, better at sustained work on long-running tasks, and more reliable at following complex instructions. For the next week, we’re doubling the included usage of the model.
#506Michael Truell@MNTRUELLComposer 2.5 is now the most-chosen model in Cursor.
We're giving everyone 10x usage for the rest of the day. Enjoy!
Cursor@cursor_aiIntroducing Composer 2.5, our most powerful model yet. It's more intelligent, better at sustained work on long-running tasks, and more reliable at following complex instructions. For the next week, we’re doubling the included usage of the model.
#583Niklas Muennighoff@MUENNIGHOFFComposer 2.5 sits on the Pareto frontier
Cursor@cursor_aiIntroducing Composer 2.5, our most powerful model yet. It's more intelligent, better at sustained work on long-running tasks, and more reliable at following complex instructions. For the next week, we’re doubling the included usage of the model.
#612Ravid Shwartz Ziv@ZIV_RAVIDSelf distillation everywhere 🥳
Cursor@cursor_aiIntroducing Composer 2.5, our most powerful model yet. It's more intelligent, better at sustained work on long-running tasks, and more reliable at following complex instructions. For the next week, we’re doubling the included usage of the model.
#694Dan Fu@REALDANFU🎼2⃣5⃣
Together AI@togethercomputeCongrats to the @cursor_ai team on Composer 2.5 — a huge milestone for agentic coding models. Together AI, the AI Native Cloud, is proud to partner on this launch. Composer 2.5 is pushing the frontier for coding agents and turning heads for its speed and quality. Excited to keep building with the Cursor team!
#716elie@ELIEBAKOUCHCursorBench vs Artificial Analysis Coding index
cursor harness seems to almost always improve score for opus, but makes the cost/task higher
opposite for gpt5.5 (codex scores are higher but cost/task with cursor is lower)
#716elie@ELIEBAKOUCHwe don't have the data point for model without cursor harness so i just put the same score, but the reality is likely lower. also very small number of data point not sure we can do some strong conclusion here but sharing anyways!
elie@eliebakouchCursorBench vs Artificial Analysis Coding index cursor harness seems to almost always improve score for opus, but makes the cost/task higher opposite for gpt5.5 (codex scores are higher but cost/task with cursor is lower)
#716elie@ELIEBAKOUCHv1 is here
elie@eliebakouchcorrelation between CursorBench and Artificial Analysis reported scores benchmarks like IFBench or tau2 show ~0 correlation with CursorBench. opus 4.7 (max effort) performs relatively better on CursorBench than on other benchmarks, gpt 5.5 shows the opposite pattern
#716elie@ELIEBAKOUCH@srush_nlp oh i need to update the plot wait
Sasha Rush@srush_nlpHuh, so you’re saying it’s smart, cheap, and fast.
#716elie@ELIEBAKOUCH@srush_nlp here you go, but honeslty not enough data point to make strong conclusion imo, just cool to see this plot imo
elie@eliebakouchCursorBench vs Artificial Analysis Coding index cursor harness seems to almost always improve score for opus, but makes the cost/task higher opposite for gpt5.5 (codex scores are higher but cost/task with cursor is lower)
#716elie@ELIEBAKOUCH@srush_nlp really cool work congrats to the team. if you can answer, do you have rough estimate on how the compute was allocated between RL and continual pt in composer 2 -> 2.5?
Sasha Rush@srush_nlpBeen working on text feedback / OPSD in Composer. Really interesting space, and a much more to be explored.
#716elie@ELIEBAKOUCHcursor is at frontier scale, both in terms of performance and compute
if composer 2.5's budget was put into a pre-train: ~6.3T total, 200B active trained on ~56T tokens
if composer 3 allocates 50% of the budget to pre-training: ~500B active, 15.3T total trained on 135T tokens.
assumptions are a lower bound: 35% MFU, FP8, ~3-4% sparsity like K2, H100 efficiency. model/token allocation is the mean between K2+K2.5 data point and Inclusion AI compute optimal rules for MoE
really impressed by the progression between composer 2 and composer 2.5
Cursor@cursor_aiIntroducing Composer 2.5, our most powerful model yet. It's more intelligent, better at sustained work on long-running tasks, and more reliable at following complex instructions. For the next week, we’re doubling the included usage of the model.
#716elie@ELIEBAKOUCH> if composer 3 allocates 50% of the budget to pre-training: ~500B active, 15.3T total trained on 135T tokens.
i meant pre/mid/continual training here (whatever is not RL), and also put 10% estimation in the attached picture
elie@eliebakouchcursor is at frontier scale, both in terms of performance and compute if composer 2.5's budget was put into a pre-train: ~6.3T total, 200B active trained on ~56T tokens if composer 3 allocates 50% of the budget to pre-training: ~500B active, 15.3T total trained on 135T tokens. assumptions are a lower bound: 35% MFU, FP8, ~3-4% sparsity like K2, H100 efficiency. model/token allocation is the mean between K2+K2.5 data point and Inclusion AI compute optimal rules for MoE really impressed by the progression between composer 2 and composer 2.5
#716elie@ELIEBAKOUCHwhen you do continual pre training at this scale on traces that look like RL rollouts, does it hurt RL if the mid training data is very similar to what you RL on? what if it's the same data but with different rollouts from another model?
intuitively i'd say yes, same intuition as what minimax M1 found with SFT cold start/RL overlap, but does this change at ~T token scale?
Cursor@cursor_aiIntroducing Composer 2.5, our most powerful model yet. It's more intelligent, better at sustained work on long-running tasks, and more reliable at following complex instructions. For the next week, we’re doubling the included usage of the model.
#716elie@ELIEBAKOUCHmaybe cc @agarwl_ @_lewtun if you have some thought here 👀
elie@eliebakouchwhen you do continual pre training at this scale on traces that look like RL rollouts, does it hurt RL if the mid training data is very similar to what you RL on? what if it's the same data but with different rollouts from another model? intuitively i'd say yes, same intuition as what minimax M1 found with SFT cold start/RL overlap, but does this change at ~T token scale?
#839Beff (e/acc)@BEFFJEZOSThis is very bullish for SpaceXAI
@elonmusk Back in the top 3 🔥🔥
Elon Musk@elonmuskTry Composer 2.5
@ericzakariasson looks good
eric zakariasson@ericzakariassoncomposer 1 was fast composer 2 was fast and intelligent composer N:
#978Jacob Jackson@JBFJAOur new model is out. It stacks up nicely against the frontier!
Cursor@cursor_aiIntroducing Composer 2.5, our most powerful model yet. It's more intelligent, better at sustained work on long-running tasks, and more reliable at following complex instructions. For the next week, we’re doubling the included usage of the model.
#980Lisan al Gaib@SCALING01yeah that's pretty good
xAI might be able to cook with Cursor data + 10T model
Cursor@cursor_aiIntroducing Composer 2.5, our most powerful model yet. It's more intelligent, better at sustained work on long-running tasks, and more reliable at following complex instructions. For the next week, we’re doubling the included usage of the model.
#1086Ashvin Nair@ASHVINAIRWe just trained another really really good model, please try it! Frontier intelligence + very fast
#1245Alex Volkov@ALTRYNENew Composer from Cursor team! Great to see their ack. to the Kimi base + how much they moved the model forward!
This isn't the one they are training on XAI Colossus, that one is coming and would likely slap hard!
Cursor@cursor_aiIntroducing Composer 2.5, our most powerful model yet. It's more intelligent, better at sustained work on long-running tasks, and more reliable at following complex instructions. For the next week, we’re doubling the included usage of the model.
#1262samsja@SAMSJA19@srush_nlp Congrats guys
Sasha Rush@srush_nlpHuh, so you’re saying it’s smart, cheap, and fast.
#1284Kevin Frans@KVFRANSWe used a pretty cool "RL with text feedback" formulation to train this one (see blog post for some details). As RL tasks get longer in horizon, I think it's a ripe time to think about ways we can extract signals that avoid the variance explosion.
Cursor@cursor_aiIntroducing Composer 2.5, our most powerful model yet. It's more intelligent, better at sustained work on long-running tasks, and more reliable at following complex instructions. For the next week, we’re doubling the included usage of the model.
#1318Sam Whitmore@SJWHITMOREcomposer 2.5 is really really great. I had it on last week for some testing, forgot that it was on, & totally didn’t realize I wasn’t on gpt 5.5 (my usual) for a while. the team did a fantastic job!!
Cursor@cursor_aiIntroducing Composer 2.5, our most powerful model yet. It's more intelligent, better at sustained work on long-running tasks, and more reliable at following complex instructions. For the next week, we’re doubling the included usage of the model.
@mntruell Congrats!
Michael Truell@mntruellComposer 2.5 is a significant step up from Composer 2. This is the very start of our work with SpaceXAI. Hope to have more improvements out soon.
#1496Chubby♨️@KIMMONISMUSIntelligence too cheap to meter. This is the real deal. Composer 2.5 is an efficiency-beast
Chubby♨️@kimmonismusHuge, did NOT expect that release. Evals looks very solid, significant jump compared to composer 2! But: it’s 10x more efficient than the competition. Looks really exciting. Need to try it out
#1496Chubby♨️@KIMMONISMUSHuge, did NOT expect that release. Evals looks very solid, significant jump compared to composer 2!
But: it’s 10x more efficient than the competition. Looks really exciting. Need to try it out
Cursor@cursor_aiIntroducing Composer 2.5, our most powerful model yet. It's more intelligent, better at sustained work on long-running tasks, and more reliable at following complex instructions. For the next week, we’re doubling the included usage of the model.
it’s rather close already ^^
🍓🍓🍓@iruletheworldmothe xai cursor combo is creating the first real challenge to openai and anthropic doesn’t hurt that anthropic are essentially giving their direct competitors billions each day expect composer to become sota quite soon.
#1733Ryo Lu@RYOLU_frontier smart extremely efficient Composer 2.5 is here
Cursor@cursor_aiIntroducing Composer 2.5, our most powerful model yet. It's more intelligent, better at sustained work on long-running tasks, and more reliable at following complex instructions. For the next week, we’re doubling the included usage of the model.
#1749Sualeh Asif@SUALEHASIF996We've gotten really really good at RL. Composer 2.5 is fighting well-above its weight class.
Very excited for the next release as we scale model sizes and FLOPs with @SpaceXAI!
Cursor@cursor_aiIntroducing Composer 2.5, our most powerful model yet. It's more intelligent, better at sustained work on long-running tasks, and more reliable at following complex instructions. For the next week, we’re doubling the included usage of the model.
@ericzakariasson Fat slow and genius
eric zakariasson@ericzakariassoncomposer 1 was fast composer 2 was fast and intelligent composer N:
#1965eric zakariasson@ERICZAKARIASSONi wrote a guide on optimizing context usage 6 months ago that i never posted. back then with the models available, you could only pick 2 of 3:
1. intelligent 2. fast 3. cheap
intelligent + fast = expensive fast + cheap = dumb cheap + intelligent = slow
now, with composer 2.5, this is no longer true and the post is obsolete. looking at TPS, avg cost per task, and score from cursorbench, it's clearly capable of all three
but benchmarks are just benchmarks. what matters is how it feels to use and if it can actually accomplish your tasks. from the feedback so far, that's very much the case
go try it out if you haven't already
@VictorTaelin just going to leave this here
Taelin@VictorTaelin@ericzakariasson I'd love to try it but unfortunately I had to sell my eyes to cover Anthropic bills, so now I can only use textual models via an API. Please let me know if you support this!
#1965eric zakariasson@ERICZAKARIASSONcomposer 1 was fast composer 2 was fast and intelligent composer N:
Cursor@cursor_aiIntroducing Composer 2.5, our most powerful model yet. It's more intelligent, better at sustained work on long-running tasks, and more reliable at following complex instructions. For the next week, we’re doubling the included usage of the model.
