Frontier Model Release Velocity Index 2026 Q2 Report

What FMRVI Measures

FMRVI is a rolling index. It counts substantive frontier releases per week across a tracked roster of labs and reports a weekly average and quarterly total. The point is not to pick a winner. The point is to tell agencies how often the ground is moving so they can size their evaluation budget and cycle accordingly.

Substantive Release Rules

A model counts as a substantive release if it clears at least one of four thresholds:

• Benchmark leadership: takes #1 or ties #1 on a widely used benchmark (SWE-bench family, MMMLU, Humanity's Last Exam, LMSYS Arena) for its capability class.
• Pricing-tier shift: establishes a new cost floor for a given capability band, for example Qwen 3.5 Flash landing at $0.065 / $0.26 per 1M for 1M-context use.
• Modality expansion: adds a meaningful new modality or context window tier (omnimodal, 1M context, vision-only variant).
• Production safeguard change: ships a new alignment, cyber, or agentic-autonomy capability that changes what the model can be deployed for at enterprise scale.

Incremental version bumps, UI changes, and product-layer launches (Claude for Excel, Computer Use on macOS, Channels) do not count as model releases even when they ship the same week. We track them separately on a product-velocity ledger because they affect client onboarding costs rather than procurement math.

Provider Cadence: Top 5 Fastest Shippers Q1 2026

The top five labs by Q1 2026 release count, ranked:

Two patterns stand out. First, Chinese labs dominate the cadence column: Alibaba, Xiaomi, and MiniMax together account for 12 of the top-5 table's 14 Q1 releases. Second, Anthropic and OpenAI appear lean on this axis but compensate with product-layer velocity. FMRVI is a model-cadence index, so those do not count here.

Q1 2026 Release Timeline

A month-by-month view of the substantive frontier releases captured by FMRVI. All dates are public launch dates, not announce dates.

The density of mid-to-late March is the single most important pattern here. Five labs shipped substantive releases in a 13-day window (March 10 through March 23). That is the point where "frontier release" stopped being a monthly headline and became a weekly one.

For a Chinese-provider-specific cut of this data, see our Chinese AI models market share report for Q2 2026.

Release-Cycle Compression

From 2023 into mid-2025, frontier model releases followed a roughly 6-month cadence per lab. OpenAI shipped GPT-4 in March 2023 and the next major capability release (GPT-4 Turbo) in November. Anthropic followed a similar pattern. Agencies could plan model evaluations on a half-year horizon with confidence that the primary vendor choice would hold.

That pattern broke during H2 2025 and collapsed outright in Q1 2026. What drove the compression:

For agencies, the practical trajectory looks like this: 6-month cycle in 2024, 3-month cycle through 2025, 4-week cycle as of Q2 2026. There is no floor yet. Daily release cadence is not coming (training cycles remain the binding constraint), but sub-4-week cadence is entirely plausible by Q4 2026.

Procurement Implications

A 4-week cycle changes the economics of agency AI procurement in concrete ways. Three shifts matter most.

1. Evaluation Cost Becomes a Structural Line Item

At 6-month cadence, model evaluation was an occasional project. At 4-week cadence, it becomes a continuous operation: you need a standing eval pipeline running against every substantive release, or you forfeit the ability to respond to pricing and capability shifts in time to matter. Budget 3 to 5 percent of total AI spend for eval infrastructure and the engineering time to run it.

2. Vendor Contracts Need Swap-In Clauses

Long-term vendor lock on a specific model is now a liability. When negotiating with cloud providers, API aggregators, or enterprise model vendors, insist on swap-in clauses that let you rotate to materially better options within 30 to 60 days without contract penalty. OpenRouter-style aggregation helps but is not a complete substitute for negotiated flexibility.

3. Spend Reserves Beat Long Forecasts

Hold 10 to 15 percent of annual AI spend as an uncommitted reserve. When a release like MiMo V2 Pro or MiniMax M2.7 resets the efficient frontier mid-quarter, the agencies that can move budget in two weeks outperform those on rigid annual plans.

For deeper data on which models currently lead the price-capability tradeoff, our efficient-frontier analysis for Q2 2026 maps the current Pareto frontier across coding, reasoning, and agentic benchmarks. For a pricing-only view, see our LLM API pricing index for Q2 2026.

FMRVI Projections for Q2

Projecting the rest of Q2 2026 is a matter of combining known backlog with observed cadence. Our base case, our upside case, and our downside case:

Seasonality Signals

Two seasonal patterns show up in the historical FMRVI data. First, early-quarter releases cluster around major conferences (Google I/O in May, OpenAI DevDay timing). Second, release density drops in the final two weeks of each quarter as labs hold back to influence quarterly earnings commentary. For Q2 specifically, expect the hottest weeks to fall between mid-May and mid-June.

Named Candidates on the Backlog

• DeepSeek V4: expected April, ~1T params, 1M context, reportedly Huawei-trained. The single most consequential pending release for open-weight benchmarking.
• Anthropic Mythos-class release: Project Glasswing framework implies a Mythos-tier production model (plausibly Claude Opus 4.7) lands in Q2 with cyber safeguards shipped.
• Gemini 3.5 line: Google's counter to GPT-5.4 native computer use is overdue; a Flash/Pro refresh pair is likely before Google I/O.
• Qwen 3.6 full line: Plus shipped April 2, with Max and Coder variants normally following within 4 weeks.

How to Build an Eval Pipeline That Keeps Up

At 4-week cadence, ad-hoc evaluation fails: by the time a senior engineer finishes benchmarking one release, two more have shipped. The eval pipeline we run for Digital Applied clients follows a specific shape.

1. A Canonical Task Set

Maintain a versioned corpus of 30 to 50 representative client tasks covering the workloads your agency actually bills for: content briefs, SEO tasks, code reviews, agentic customer workflows, data extraction, dashboard generation. Each task has a rubric, a reference answer, and a token budget. This is the single largest investment in the pipeline and pays back every time a new model ships.

2. Automated Trigger on Release Events

Subscribe to model-announcement feeds from the top ten labs, OpenRouter's new-model webhook, and Hugging Face trending releases. Any substantive release triggers a run of the canonical task set within 24 hours. No human in the loop for the initial score; humans only review the delta.

3. Three-Dimensional Scoring

Rate each model on quality, cost, and latency per task, then aggregate into a per-workload Pareto chart. A model that beats the current default on quality at higher cost may still lose for bulk workloads and win for high-stakes work. One score is never enough.

4. Production Shadow Mode

When a candidate clears the canonical evals, shadow-deploy it against 1 to 5 percent of live traffic for 48 hours before making any routing decisions. Production behavior often diverges from benchmark behavior in ways that only real traffic exposes.

FMRVI vs Spend and Volume

Does release velocity correlate with adoption? The short answer is not in the way you might expect. FMRVI rank, intelligence rank, and usage rank diverge sharply.

Taking April 2026 OpenRouter data at face value: MiMo V2 Pro is #1 in weekly tokens (4.79T) and #1 in coding tokens (2.05T) but #10 in the Artificial Analysis Intelligence Index. GPT-5.4 is #1 in intelligence but far from #1 in usage. Claude Opus 4.6 is #6-#7 in usage but a clear leader in production agentic coding. The three rankings measure different things, and an agency optimizing on the wrong one picks the wrong model.

What FMRVI Correlates With

• OpenRouter share gains: high-cadence labs (Xiaomi, Alibaba) are gaining share fastest. Correlation is strong because rapid iteration lets them match price and capability moves quickly.
• Enterprise adoption: high-cadence labs correlate negatively with enterprise share in Q1 2026. Enterprise buyers want stability, and a model that ships a successor every 6 weeks looks risky.
• Coding workload share: modest positive correlation. Coding-tool vendors (Cursor, Windsurf, Claude Code) move fast and benefit from new options.
• Intelligence-benchmark rank: no meaningful correlation. Many high-cadence labs skip expensive benchmarking, and many high-intelligence models release slowly.

For a granular view of the current OpenRouter rankings driving the adoption side of this picture, see our OpenRouter rankings for April 2026. For a deeper look at how MiniMax M2.7 reshaped the agentic coding tier in particular, see our MiniMax M2.7 agentic coding release guide.

What FMRVI Does Not Capture

Every index has blind spots, and FMRVI is no exception. Four limitations to understand before citing the number.

1. Closed-Source Partner Pilots

Labs routinely ship pre-GA partner builds to selected enterprise customers weeks before public release. Those pilots influence the market (Cursor Composer 2 was built on Kimi K2.5 before K2.5 was widely available, for example) but do not count in FMRVI until they hit public availability.

2. In-Product Model Swaps

ChatGPT, Gemini, Copilot, and Claude consumer products frequently route requests to different models behind the scenes. A silent backend swap that moves all ChatGPT traffic to GPT-5.4 is commercially enormous but is not a release event.

3. Soft or Incremental Updates

Gemini 3.1 Flash-Lite (Mar 3) or Grok 4.20 beta (Feb 17) are real releases but borderline on our substantive threshold. We include them only when they meet one of the four rules, conservatively.

4. Region-Locked Launches

Some releases are initially available in one region (China-first, EU-first) before rolling out globally. FMRVI uses the first-region release date, but that can understate the effective cadence for any particular market.

Conclusion

The Frontier Model Release Velocity Index makes one uncomfortable fact concrete: the agencies that were planning AI procurement on a 6-month horizon in 2025 are now shipping into a 4-week market. Our Q1 2026 count of 12+ substantive frontier releases is double Q4 2025, and our base case for Q2 is 14 to 18 more.

The response is not to chase every release. It is to build the evaluation, procurement, and portfolio infrastructure that can absorb new releases without blowing up client delivery plans. A canonical eval corpus, automated release triggers, three-dimensional scoring, and shadow deployment are the minimum. For related analysis on the market structure driving these releases, our open-weight vs closed-source AI models analysis for Q2 2026 covers the strategic split that makes this cadence sustainable.

Frequently Asked Questions