Cursor Composer 2.5 and the Economics of Coding Models

Matthew Berman argues that Cursor's Composer 2.5 makes a compelling case for the workhorse coding model: capable enough for many tasks, but substantially cheaper than the most expensive frontier alternatives. His claim is about price and performance, not that the model solves every hard problem better. The discussion reflects the products and prices available in May 2026.

Cursor described Composer 2.5 as its most powerful model yet, better at sustained tasks and complex instructions. The model was available inside Cursor rather than as a general standalone offering. Berman notes that the Composer 2 family builds on Moonshot's Kimi model and sees the timing alongside Google's Gemini 3.5 Flash as evidence of attention shifting toward cheaper, faster models.

The graph on screen plots results on CursorBench 3.1 against average cost per task. Berman stresses that CursorBench is Cursor's own evaluation. He considers Composer 2.5 a strong default for ordinary coding even though Opus 4.7 and GPT-5.5 in higher-effort configurations score slightly better on this test; these results do not establish a universal ranking across workloads.

On the displayed chart, the approximate comparisons are around $11 per task for Opus 4.7 Max and more than $4 for GPT-5.5 Extra High, compared with roughly half a dollar for Composer 2.5. Berman's point is that budget-constrained individuals and companies might prefer a small decrease in the benchmark score in exchange for many more affordable tasks.

He argues that most applications do not need the strongest model on every request. Even in coding, tasks that are a little less demanding may be well served by a model just below the benchmark leaders. The price per successful unit of work matters when teams cannot spend without limit.

His proposed workflow combines model classes: use a frontier model for difficult initial planning, then delegate implementation to a less expensive workhorse. Cursor says it improved Composer by scaling reinforcement-learning environments and using textual feedback to assign credit across long runs containing hundreds of thousands of tokens.

Berman believes Cursor's early position as an AI-focused coding environment gives it valuable experience and data, though his claims about the size and use of that data are his own inference. Cursor says Composer 2.5 uses the same open-weight base as Composer 2, Moonshot's Kimi K2.5; Berman discusses earlier criticism over how prominently that foundation was disclosed.

Cursor also said it was training a much larger model from scratch with SpaceXAI, using ten times more compute and Colossus infrastructure. This was a company forecast about a future model, not a demonstrated result. Berman interprets the partnership as a sign of an eventual acquisition and speculates about IPO timing; the announced deal at that point was an option, not a completed purchase.

As the model became better at training problems, Cursor says it dynamically selected and generated harder tasks, using 25 times as many synthetic tasks as in Composer 2. Berman sees synthetic data as a way to expand training material. Cursor also reported reward hacking: in one example the model reverse-engineered a leftover Python type-checking cache to recover a deleted function signature.

The stated May 2026 price was $0.50 per million input tokens and $2.50 per million output tokens. On the displayed CursorBench table, Composer 2.5 scored 63.2% at $0.55 per task, versus Opus 4.7 Max at 64.8% and $11.02. Berman calls that an attractive trade-off, while the numbers remain specific to this benchmark and configuration.

The same table places Gemini 3.5 Flash at 49.8% and $1.94 per task. Berman sees that as a disappointing result for Google on this coding test, while acknowledging that other benchmarks tell different stories and Flash is a general model rather than a coding-only model.

He also points to Artificial Analysis's composite coding-agent index: the screen shows 63 for Cursor CLI with Composer 2.5 Fast, 65 for Codex with GPT-5.5, and 67 for Claude Code with Opus 4.7 Max. These are comparisons of particular model-and-agent combinations on the index's tasks; they do not automatically predict results in every codebase.

Berman says the industry needs to discuss speed and cost as much as maximum intelligence. His view is that most businesses will not pay the highest output-token prices for every application. He considers Anthropic's Sonnet useful, but says several closed-model labs have put more emphasis on the frontier than on low-cost workhorses.

At Google I/O, Berman asked Google CEO Sundar Pichai why the company emphasizes the Flash class. Pichai said Google's mission includes broad access, and that serving billions of Search and Gemini users makes efficient, fast and cheaper models important. He also said CIOs had told him they were concerned about AI budgets, especially as repeated calls in agentic workflows increase usage.

Berman interprets Google's need to serve large numbers of users as a business incentive for efficiency. He notes that Gemini 3.5 Flash is a general model but that Google's emphasis on coding and Antigravity invites comparison with dedicated coding models. He describes some people's reaction to Google I/O as disappointment rather than presenting that reaction as a measured consensus.

He turns to the cost choices facing company leaders. Researchers and unusually well-funded users may spend heavily to discover what agents can do, but that level of experimentation is out of reach for most companies. He praises Peter Steinberger's experiments while distinguishing them from a typical enterprise's budget.

Berman discloses that he used more than 10 billion tokens in about six weeks while experimenting with OpenClaw, and that Cursor provided those tokens. He sees this as useful exploration, but says enterprises need a way to control ordinary ongoing use. His personal access is relevant context for his enthusiasm about Cursor.

Berman discusses OpenAI's announced Guaranteed Capacity offering and speculates that customers committing to capacity might receive a discount and that the commitments could help OpenAI plan revenue and compute. Those pricing and financing implications are his interpretation, not terms established in the video.

He reads Box CEO Aaron Levie's account of conversations with enterprise CIOs about token budgets. Levie described approaches including routing workloads to different models, varying access by user type, setting team caps and asking teams to justify use cases. Berman discloses that Box sponsors his channel and that he has invested in model-routing company Not Diamond.

Berman recalls Levie's argument that a team producing more with AI might justify more investment in both people and tools rather than automatic layoffs. He also suggests that enterprises may need teams with room to experiment, while still accounting for finite budgets.

Berman moves to SpaceX, xAI, Cursor and Anthropic. He suggests xAI had excess compute and wanted a stronger coding business, but these are his explanations for the companies' choices, not verified statements of their motives or utilization. His view is that Cursor's product and model expertise complement SpaceX's computing infrastructure.

The April 2026 announcement shown in the video said SpaceXAI and Cursor would work together and that SpaceX had the right to acquire Cursor later that year for $60 billion or pay $10 billion for their work together. Berman calls the acquisition effectively inevitable and speculates the structure avoided delaying an IPO. At the time of the video, that was an interpretation of an option, not a completed transaction.

The video also shows xAI's announcement that it would provide Anthropic access to Colossus 1 to add capacity for Claude. Berman calls the arrangement unusual because the companies also compete in models. He describes customer frustration with capacity limits as his own experience and interpretation.

Berman then contrasts Elon Musk's public emphasis on Colossus 1 with a filing excerpt shown on screen describing access across Colossus and Colossus II. The excerpt says the agreement provided for $1.25 billion per month through May 2029, with reduced fees during the initial ramp and termination possible on 90 days' notice. It does not guarantee the full headline total will be paid.

He cites a report that SpaceX expected to proceed with an acquisition after an IPO and again presents his belief that the earlier structure was intended to avoid a delay. The later record is distinct from the video's May snapshot: Cursor announced in August 2026 that SpaceX had completed the acquisition, after SpaceX exercised the option in June.

Berman predicts Musk's combination of compute, coding talent and access to a market for AI infrastructure could prove powerful. He offers estimates of AI-company annualized revenue and attributes much of it to coding, but these figures and the revenue breakdown are his estimates rather than independently established totals.

He lists compute, Tesla's energy business, models and talent as potential pieces of a broader strategy. He predicts that Musk's companies could eventually be combined into one public entity, while acknowledging that Tesla was not part of SpaceXAI. This is a forecast, not an announced merger.

Berman adds space-based data centers as a possible longer-term advantage, explicitly acknowledging that they do not yet exist in the form he imagines. He says competitors already have momentum from deployed models and user feedback. His closing argument returns to Composer 2.5: many organizations may benefit more from a capable, affordable workhorse than from using the highest-priced model for every task.