LMC Capability Drift Tracker
How LLM capabilities spread through the catalog, quarter by quarter. Each curve shows the percentage of models released in a given release cohort that shipped with a given capability. 466 models tracked across 12 release quarters.
Catalog last refreshed 2026-10-12 (auto-updated hourly from live provider APIs).
Capability Adoption Curves
Percentage of models released in each quarter that ship with the given capability.
Cohort Capability Matrix
Share of each release-quarter cohort that ships each capability, expressed as a percentage.
| Cohort | N | Reasoning | Vision | Function Calling | JSON Mode | Web Search | Image Output |
|---|---|---|---|---|---|---|---|
| 2024 Q1 | 2 | 0% | 0% | 100% | 100% | 0% | 0% |
| 2024 Q2 | 11 | 0% | 45% | 55% | 64% | 18% | 18% |
| 2024 Q3 | 19 | 0% | 21% | 53% | 63% | 5% | 21% |
| 2024 Q4 | 20 | 15% | 20% | 55% | 40% | 5% | 10% |
| 2025 Q1 | 25 | 32% | 40% | 40% | 68% | 32% | 0% |
| 2025 Q2 | 30 | 60% | 77% | 93% | 83% | 60% | 0% |
| 2025 Q3 | 43 | 63% | 40% | 91% | 84% | 28% | 0% |
| 2025 Q4 | 50 | 70% | 74% | 86% | 90% | 46% | 8% |
| 2026 Q1 | 55 | 85% | 65% | 89% | 89% | 38% | 2% |
| 2026 Q2 | 66 | 98% | 83% | 91% | 89% | 48% | 6% |
| 2026 Q3 | 130 | 95% | 72% | 95% | 89% | 46% | 0% |
| 2026 Q4 | 8 | 88% | 75% | 100% | 75% | 38% | 13% |
Capability Debuts
First model in the catalog to ship each capability and how far the capability has spread since.
| Capability | First model | Latest adoption |
|---|---|---|
Reasoning | Falcon3 10B Instruct | 88% |
Vision | GPT-4 Turbo | 75% |
Function Calling | GPT-3.5 Turbo | 100% |
JSON Mode | GPT-3.5 Turbo | 75% |
Web Search | GPT-3.5 Turbo (batch) | 38% |
Image Output | DALL-E 3 | 13% |
Latest models shipping reasoning
The fastest-spreading capability over the last year is reasoning. Here is what recently shipped with it.
| Model | Provider | Released |
|---|---|---|
| Step 5 Preview | StepFun | 2026-10-08 |
| Claude Haiku 5.5 | Anthropic | 2026-10-07 |
| Claude Haiku 5.5 (batch) | Anthropic | 2026-10-07 |
| Nano Banana 2.1 | 2026-10-06 | |
| Mistral Large 4 | Mistral AI | 2026-10-06 |
| Ling 3.1 Flash | inclusionai | 2026-10-02 |
| Apodex 1.1 Mini (free) | apodex | 2026-10-01 |
| GPT-6.1 Sol Pro | OpenAI | 2026-09-29 |
| GPT-6.1 Sol Pro (batch) | OpenAI | 2026-09-29 |
| GPT-6.1 Sol | OpenAI | 2026-09-29 |
| GPT-6.1 Sol (batch) | OpenAI | 2026-09-29 |
| Claude Sonnet 5.5 | Anthropic | 2026-09-28 |
How the Capability Drift Tracker works
Every model we index has a release timestamp and a binary capability manifest. We bucket models into release-quarter cohorts using that timestamp, then compute the share of each cohort that shipped with a given capability. The result is an adoption curve per capability per quarter.
We exclude cohorts before 2024 Q1 from the visual curves because those quarters carry fewer than a dozen tracked releases each, so a single model can swing the percentage by tens of points. The cohort table below the charts shows all included quarters and their sample sizes.
Capability flags follow provider-declared manifests reconciled against documented API surfaces. A model that can accept images is counted under vision. A model that exposes a tool-calling endpoint is counted under function calling. A model that supports an explicit reasoning mode (visible or hidden thinking tokens) is counted under reasoning, even if the base mode is plain chat.
Related trackers
It is a quarter-by-quarter view of how the LLM catalog has adopted each of six core capabilities: reasoning, vision, function calling, JSON mode, web search, and image output. Each cohort is defined by release quarter, and every model in that cohort contributes to the capability percentages for that quarter. This lets us see which capabilities are becoming table-stakes and which are still rare. Across the latest cohort (2026 Q4, N=8), 88% of new releases shipped with reasoning support and 75% shipped with vision input.
Provider roll-up hides the temporal story. A provider that shipped a reasoning model in 2024 Q4 and a non-reasoning model in 2025 Q2 would average out to 50% if we grouped by provider. Release-cohort grouping lets us see the calendar-quarter at which each capability crossed 50% adoption, which is a much more actionable question for buyers planning a rebuild.
Reasoning is the fastest-spreading capability over the last year. It was at 70% one year ago and now sits at 88% of new releases, a gain of 18 percentage points. Extended chain-of-thought / deliberate reasoning mode with visible or hidden thinking tokens.
Yes. The denominator for each cohort is every tracked model released in that quarter, including open source, free, and paid-only. We do not filter by license or pricing because capability adoption is a platform-wide question, and excluding open source would understate how quickly the open ecosystem is catching up on things like reasoning and vision.