Modelglass Pricing Index — 2026 Q3
Window: 2026-06-16 → 2026-09-14 (trailing 90 days) · Published 2026-09-15
Every AI model pricing dataset goes stale the moment it's published — this one doesn't try to be a snapshot. Modelglass tracks pricing, launches, retirements, and benchmark scores continuously across image, language, video, audio, and 3D-generation models; this report is a quarterly read of what actually moved, pulled directly from that live registry rather than assembled from vendor announcements after the fact.
This first edition uses a trailing-90-day window rather than a strict calendar-quarter comparison: Modelglass has been tracking this data continuously for only about three months, so a Q3-vs-Q2 comparison would either compare against a period before consistent tracking existed or silently exclude models added mid-quarter. Every figure below is honest about its own coverage on that basis — the 2026 Q4 edition will be the first genuine quarter-over-quarter comparison.
Every price figure carries a source URL and a verified-at date on the underlying registry entry — this report adds no unsourced numbers, and sources are cited inline per finding rather than only linked at the bottom.
Price changes
9 offerings across 2 of the 5 tracked modalities, language and audio, had a price change in the window: the same offering repricing, not a new listing. Image, video and 3D had none. 3D is the youngest vertical, launched inside this same window, so nothing in it has a second price point yet.
| Model | Modality | Change | Δ | Effective | Source |
|---|---|---|---|---|---|
| OpenAI o3 | LLM | $10.00 → $2.00 /1M input tokens | -80% | 2026-08-29 | platform.openai.com/docs/pricing |
| OpenAI o3 | LLM | $40.00 → $8.00 /1M output tokens | -80% | 2026-08-29 | platform.openai.com/docs/pricing |
| Qwen 2.5 72B (Together AI) | LLM | $0.36 → $1.20 /1M input tokens | +233% | 2026-07-22 | together.ai/models/qwen-2-5 |
| Qwen 2.5 72B (Together AI) | LLM | $0.40 → $1.20 /1M output tokens | +200% | 2026-07-22 | together.ai/models/qwen-2-5 |
| Llama 4 Maverick (Together AI) | LLM | $0.15 → $0.27 /1M input tokens | +80% | 2026-07-22 | together.ai/models/llama-4-maverick |
| Llama 4 Maverick (Together AI) | LLM | $0.60 → $0.85 /1M output tokens | +42% | 2026-07-22 | together.ai/models/llama-4-maverick |
| Llama 4 Scout (Together AI) | LLM | $0.10 → $0.18 /1M input tokens | +80% | 2026-07-22 | together.ai/models/llama-4-scout |
| Llama 4 Scout (Together AI) | LLM | $0.30 → $0.59 /1M output tokens | +97% | 2026-07-22 | together.ai/models/llama-4-scout |
| DeepSeek V4-Flash | LLM | $0.14 → $0.44 /1M input tokens (peak-hour rate) | +214% | 2026-08-16 | api-docs.deepseek.com (archived) |
| DeepSeek V4-Flash | LLM | $0.28 → $1.32 /1M output tokens (peak-hour rate) | +371% | 2026-08-16 | api-docs.deepseek.com (archived) |
| DeepSeek V4-Pro | LLM | $0.435 → $1.32 /1M input tokens (peak-hour rate) | +203% | 2026-08-16 | api-docs.deepseek.com (archived) |
| DeepSeek V4-Pro | LLM | $0.87 → $3.96 /1M output tokens (peak-hour rate) | +355% | 2026-08-16 | api-docs.deepseek.com (archived) |
| ElevenLabs Flash v2.5 | Audio | $60 → $50 /1M chars | -17% | 2026-07-30 | elevenlabs.io/pricing/api |
| ElevenLabs Multilingual v2 | Audio | $120 → $100 /1M chars | -17% | 2026-07-30 | elevenlabs.io/pricing/api |
| Azure Neural TTS | Audio | $16 → $15 /1M chars | -6% | 2026-08-31 | azure.microsoft.com/pricing |
The biggest cut this window is OpenAI taking 80% off o3 in both directions ($10/$40 → $2/$8 per 1M tokens), a frontier reasoning model repricing into a materially cheaper tier. The biggest rises are DeepSeek V4-Flash and V4-Pro, up 203–371% at peak hours when DeepSeek's peak and off-peak pricing started on 16 Aug. The Together AI open-weight rehosts (Qwen 2.5 72B, Llama 4 Maverick/Scout) also moved up, by 42–233%: a reminder that "open-weights" doesn't mean "price-stable" when the number that moves is the hosting price, not a licensing fee.
New listings this window
63 distinct offerings were priced for the first time in the 90-day window, across all 5 modalities — roughly 2 new offerings added every 3 days. Highlights by modality:
- LLM: Claude Opus 5, Claude Sonnet 5, GPT-5.6 (Sol/Luna/Terra), GPT-5.5 Pro, Gemini 3.1 Pro, Gemini 3.5 Flash, Gemini 3.8 Flash, DeepSeek V4-Flash/V4-Pro, GLM-5.1/5.2, Kimi K3, ERNIE 5.1, Reka Flash, Inkling.
- Image: Seedream 4.0/4.5/5.0 Pro, Gemini 3.1 Flash Image, Gemini 3 Pro Image, Gemini 2.5 Flash Image, FLUX.1 Kontext, Luma Uni-1.1, Krea-2-Turbo.
- Video: Veo 3/3.1, Seedance 2/2.0/2.0 Fast/2.0 Mini, LTX-2.3, Vidu Q3, FLUX 3 Video, Gemini Omni Flash/1.1 Flash, Grok Imagine Video 1.5, Aleph 2.
- Audio: ElevenLabs Eleven v3, Fish Audio S1/S2/S2.1 Pro, Inworld TTS-1.5 Max, Grok Voice Think Fast 2.0, Stable Audio 2.5/3.0.
- 3D (new vertical): Meshy 7, Hunyuan3D, Hi3D, TRELLIS, Rodin, Stable Point Aware 3D, Stable Fast 3D — this vertical's entire current roster was priced inside this window, since it launched during it.
Currently retired or deprecated
14 offerings across the tracked registries are marked retired or deprecated as of 2026-09-14. This is a census, not a "retired in this window" count — status is a point-in-time field, not a dated event.
| Model | Modality | Status |
|---|---|---|
| Imagen 4 | Image | retired |
| DALL·E 3 | Image | deprecated |
| GPT Image 1 | Image | deprecated |
| DeepSeek R1 | LLM | retired |
| DeepSeek V3 | LLM | retired |
| K-EXAONE-236B-A23B | LLM | retired |
| Grok 3 | LLM | retired |
| Mochi 1 | Video | deprecated |
| Gemini Omni Flash | Video | deprecated |
| Veo 2 | Video | deprecated |
| Veo 3 | Video | deprecated |
| Gen-3 Alpha | Video | retired |
| PlayHT 2.0 | Audio | retired |
| Stable Audio 2.0 | Audio | deprecated |
Quality-adjusted cost: coding vertical
Raw $/token doesn't tell you which model is actually the best value — a cheap model that fails half the time isn't cheap. Joining LLM pricing against the coding vertical's SWE-bench Verified scores gives a cost-per-benchmark-point view: blended $/1M tokens (average of input and output) divided by SWE-bench Verified score. 28 models had both a live price and a score this window — cheapest cost-per-point (best value) shown below, trimmed to the notable rows.
| Model | SWE-bench Verified | $/1M in | $/1M out | Blended $/1M | $ per benchmark point |
|---|---|---|---|---|---|
| DeepSeek V3.2 | 73.1% | $0.28 | $0.42 | $0.35 | $0.0048 |
| MiniMax M3 | 80.5% | $0.30 | $1.20 | $0.75 | $0.0093 |
| Mistral Large 3 | 41.4% | $0.50 | $1.50 | $1.00 | $0.0242 |
| Kimi K2.5 | 70.0% | $0.60 | $3.00 | $1.80 | $0.0257 |
| Gemini 2.5 Flash | 54.0% | $0.30 | $2.50 | $1.40 | $0.0259 |
| GPT-5.6 Luna | 93.0% | $1.00 | $6.00 | $3.50 | $0.0376 |
| Claude Sonnet 5 | 85.2% | $2.00 | $10.00 | $6.00 | $0.0704 |
| Kimi K3 | 93.4% | $3.00 | $15.00 | $9.00 | $0.0964 |
| Claude Sonnet 4.6 | 79.6% | $3.00 | $15.00 | $9.00 | $0.1131 |
DeepSeek V3.2 costs about 24 times less per benchmark point than Claude Sonnet 4.6 ($0.0048 vs $0.1131), for a score 6.5 points lower (73.1% vs 79.6%). Only the coding vertical is built out in this first edition, as the methodology proof; science, agentic, and GDPval use the same join pattern against their own benchmark scores and are the natural next additions.
Methodology and sourcing
All figures were pulled live from Modelglass's own registry artifacts —
covering image, LLM, video, audio, and 3D pricing — rebuilt fresh from
the committed registry and ontology source on 2026-09-14, not read from
a stale copy. Every price carries a source.url and verified_at pair on the underlying registry
entry; two entries with a recorded prior price of exactly $0 were
excluded from the price-changes table as likely data artifacts from
initial entry rather than confirmed free-to-paid transitions, pending a
data-side recheck.
Frequently asked questions
- What is the Modelglass Pricing Index?
- A quarterly report on real, sourced AI model pricing — what changed, what launched, what retired, and a quality-adjusted cost-per-benchmark-point view — pulled fresh from Modelglass's own registry, not vendor press releases.
- How is "cost per benchmark point" calculated?
- Blended $/1M tokens (average of input and output price) divided by a model's score on an independent benchmark — SWE-bench Verified for the coding vertical in this edition. Lower is better value, not just cheaper.
- Where does the pricing data come from?
- Every figure carries a source URL and a verified-at date on the underlying registry entry, so this report adds no unsourced numbers. Sources are cited inline per row, not just linked at the bottom.
- When's the next edition?
- Quarterly, ongoing. This first edition uses a trailing-90-day window rather than a calendar quarter, since consistent tracking has only existed for ~3 months; the 2026 Q4 edition will be the first genuine quarter-over-quarter comparison.
Correction, 7 Nov 2026: this report first said 18 offerings across 4 modalities changed price. Its table also listed changes for Command R+, MiMo-V2.5-Pro and GPT-Audio 1.5 that we had recorded wrongly, and it left out DeepSeek V4-Flash and V4-Pro's 16 Aug change. The count, table and summary now match our corrected records. We also corrected the DeepSeek V3.2 and Claude Sonnet 4.6 comparison from roughly 15x to about 24 times, and the score gap from "within 7 points" to 6.5 points.