The same coding success rate can come with very different bills. Databricks launched its public AI Index on November 16, combining usage data from more than 10,000 companies with internal coding-agent tests. In one comparison, two models matched on task success while their average costs differed by roughly 32 times, according to the company.
The resource joins two distinct views of enterprise AI: what companies use through Unity Gateway, Databricks’ service for accessing and governing models, and how coding agents perform on its own engineering work. Databricks says it will publish weekly evaluation updates as it assesses new proprietary and open-weight releases.
Equal scores, unequal bills
GPT-6 Luna and Claude Opus 5.5 each achieved an 89.1% task success rate in Databricks’ benchmark. Their costs were far apart. The company’s explanation goes beyond the price of tokens, the units of text models process: extra exchanges, longer outputs and failed attempts can all increase the work needed to finish.
Average cost per coding task
01$0.05GPT-6 Luna
Databricks measured an average cost of $0.05 per task at an 89.1% success rate.
02$1.62Claude Opus 5.5
Databricks measured an average cost of $1.62 per task at an 89.1% success rate.
Gemini 3.8 Flash illustrates the difference between cheap tokens and cheap completed work. Despite its low token price, it averaged 140 turns per task and cost $11.97 per task, placing it among the benchmark’s most expensive models.
At the top of the quality ranking, Claude Fable 5.1 scored 96.0% at $7.47 per task. Muse Wonder scored 94.9% at $1.53. In those tests, the extra 1.1 percentage points cost nearly five times as much.
These are model-and-tool results, not isolated model scores. Databricks evaluates models using Claude Code and Codex—the software that supplies tools and manages an agent’s work. The index displays each model in whichever of those two setups sits closest to the best cost-quality trade-offs.
More providers, with a counting caveat
The adoption data shows customers moving beyond one supplier. Databricks says the share of Unity Gateway enterprises using only one model provider fell from about 64% in July to 28% by November 2, 2026.
That measure tracks accumulated provider use, not necessarily simultaneous use. Among companies active in a given week, it counts every provider used since July 20. It also treats Gemini and Gemma as one provider, and GPT and GPT-OSS as one.
Open-weight models also gained share, rising from 7.3% of token traffic in July to 30.9% in November, Databricks reports. The index’s proprietary-versus-open-weight traffic chart covers Databricks-hosted models on Unity Gateway. In the separate coding evaluation, DeepSeek V4.1 and GLM 5.3 tied for third with 93.1% pass rates.
What the measurements capture
The enterprise sample is aggregated and anonymized, and includes traffic routed to external providers such as Anthropic and OpenAI. The coding results come from a different source: Databricks’ own private codebase and engineering workflows, including creating and reviewing code changes and investigating issues.
Its July benchmark methodology describes how Databricks constructed those tests from recent, human-written pull requests—proposed code changes. The process applied several filters:
- Exclude bot, service-account, fully AI-generated and automatically generated changes.
- Select self-contained changes with strong tests, spanning typical work across languages and systems.
- Set aside the original tests, then use them to grade the agent’s solution rather than relying on an AI judge.
Spending figures have another important boundary. Wallet share uses providers’ published list prices and cache-read rates; it is not what Databricks bills. Weekly human-user spending likewise prices input, cached and output tokens at list rates.
On that basis, median weekly spending reached $2.49, up from $1.10 over six weeks, while the 99th percentile stood at $699.74. Typical-user consumption was rising even as spending at the 90th and 99th percentiles stayed relatively flat—a different budgeting signal from the cost of the most intensive users alone.
Reader comments
Newest comments first. Replies stay oldest first.