224 New Models in One Quarter? Only 95 Are Real: How Q3 2026's LLM Launches Are Priced
Our database logged 224 new model rows between July and September 2026. After removing batch mirrors, aliases and free duplicates, 95 distinct models remain, and seven of the eight points on the price-versus-intelligence frontier belong to them. Data as of 2026-10-05.
1. What counts as "new", and how we price it
Everything below comes from this site's own database (data as of 2026-10-05), which is fed daily from OpenRouter. "New" means a model row whose created_at falls between 2026-07-01 and 2026-09-30. That timestamp records when our scraper first saw a listing, not the day a vendor launched the model. The difference is not cosmetic: 68 of the 224 rows carry an OpenRouter creation date earlier than 1 July (61 batch mirrors, 4 base models and 3 free listings of older models), so they were never Q3 launches at all.
Prices are USD per 1M tokens, collapsed into a single number with a 3:1 input-to-output blend, (3 × input + output) / 4, a reasonable proxy for chat and RAG traffic that reads more than it writes. A single day's price is a poor input, because OpenRouter's model-level price follows whichever endpoint is currently the default, and that flaps. The current row for GLM 5.2 reads $0.024 in / $16 out; its 31-day median is $0.65 / $2.04. Kimi K3 reads $0.67 / $14 today against a median of $3.00 / $15. DeepSeek V4 Flash 0731 reads $0.0152 / $1.28 against a median of $0.04 / $0.28. So every price below is the median of daily snapshots from 2026-09-05 to 2026-10-05, with input and output taken separately and then blended. Three of the 91 new models (Claude Opus 5 Fast, KAT-Coder-Air V2.5, Qwen3.8 Max) have no snapshot in that window (none of the three is in OpenRouter's catalog any more) and use their last recorded price.
"Intelligence" is the Artificial Analysis (AA) Intelligence Index stored in our benchmarks table on 2026-09-08. Only 31 of the 91 new models carry a score, and the 32 that appeared after 8 September cannot have one yet, so the value analysis covers a third of the cohort and leans toward earlier, better-known launches. These numbers may also differ from older posts (the stored GPT-5.4 score is 39.0 in this snapshot), so compare figures within one table, not across articles.
2. From 224 rows to 95 models
The headline count is inflated by listing mechanics, not by vendors. The table removes one kind of duplicate at a time.
| Step | Rows | Left |
|---|---|---|
| Rows first seen 1 Jul – 30 Sep (Jul 61, Aug 87, Sep 76) | — | 224 |
Batch mirror listings (:batch) | −94 | 130 |
~…-latest aliases | −10 | 120 |
:free duplicates | −20 | 100 |
| Older models first seen late (Ministral 8B, Nemotron 3.5 Content Safety, Nex-N2-Mini, Venice Uncensored) | −4 | 96 |
| Routers and stealth models (Auto Router Beta, Jev Router, Ox Alpha, Union Alpha, Space Bunny Alpha) | −5 | 91 |
| Free-only models added back (Ling 3.0 Tiny, Ling 3.0 Flash Sante, LFM2.5-2.6B, Dots3-Note Preview) | +4 | 95 |
The biggest block is batch mirrors: 94 rows, 42% of the total. Seventy-six of them appeared on just three days (27 on 28 July, 32 on 6 August, 17 on 28 August), and 61 mirror models that already existed before July, from GPT-3.5 Turbo to Claude Sonnet 5. That is why Mistral shows seven rows and zero new models: six are batch mirrors and the seventh is Ministral 8B, which OpenRouter dates to October 2024.
We kept three preview or experimental builds among the 95 (DeepSeek V4 Flash Vision Exp, Tencent Hy4 preview, Mercury 2.5 Preview) because they are separate weights with their own prices, but they are lower-commitment launches than the rest. OpenAI's 14 models are really seven models plus seven "Pro" twins: in our data every Pro SKU carries the same per-token price as its base SKU. We have no data on how many reasoning tokens each one burns, so cost per task can still differ.
De-duplication also moves the peak. Raw rows say August was busiest (61, 87 and 76 for July, August and September), but 50 of August's 87 rows are batch mirrors. By distinct model the busiest month was September: 26, 26 and 43, with September alone holding 45% of the quarter.
3. Who shipped the most
| Provider | Raw rows | Distinct models | What inflated the raw count |
|---|---|---|---|
| OpenAI | 63 | 14 | 45 batch, 4 aliases |
| Anthropic | 19 | 5 | 14 batch |
| 16 | 4 | 12 batch | |
| DeepSeek | 12 | 4 | 4 batch, 3 aliases, 1 free |
| Qwen | 11 | 8 | 2 batch, 1 free |
| Z.ai | 10 | 4 | 3 batch, 2 aliases, 1 free |
| inclusionAI | 8 | 5 | 5 free listings |
| Meta | 7 | 6 | 1 batch |
| Mistral | 7 | 0 | 6 batch, 1 older model |
| Tencent | 6 | 5 | 1 free |
| All other providers | 65 | 40 | — |
| Total | 224 | 95 |
Ranked by raw rows, OpenAI (63), Anthropic (19) and Google (16) lead. By distinct models the order is OpenAI (14), Qwen (8) and Meta (6); Anthropic and Google fall to 5 and 4 because 14 and 12 of their rows are batch mirrors. The top three account for 28 of the 95 models (29%). The rest is a long tail of 32 providers in total, 11 of which shipped exactly one model. Our provider country field is empty for 19 of the 95, so we do not split by country.
By first-seen date, OpenAI's 14 arrived in four drops: the GPT-5.6 Luna, Terra and Sol family on 9 July, GPT-6 Astra on 5 September, GPT-6 Luna and Sol on 22 September, and GPT-6.1 Sol on 29 September. Meta's six are Muse Spark 1.1 to 1.3 (with two contributor tiers) and Muse Glimmer 30B. Qwen's eight are one 3.7 Flash and seven 3.8 variants.
The open-weight share fell. OpenRouter lists a Hugging Face ID for 32 of the 91 new models (35%), against 156 of 338 older ones (46%). That is a proxy, not a licence audit, but it points the same way as the top end of the market: of the 13 new models scoring 40 or more on AA, only four have downloadable weights (Kimi K3, Qwen3.8 2.4T A95B, GLM 5.3 and GLM 5.3 Flash). Compare with our earlier look at open-weight value kings.
4. The new cohort vs the older catalog: a barbell
The older baseline is every base listing first seen before July with a non-zero price (n = 338), measured the same way. All figures are USD per 1M tokens.
| Blended 3:1 and per-side prices | Q3 Q1 | Q3 median | Q3 Q3 | Older Q1 | Older median | Older Q3 |
|---|---|---|---|---|---|---|
| Input $/M | $0.095 | $0.50 | $2.00 | $0.15 | $0.40 | $1.25 |
| Output $/M | $0.345 | $2.50 | $6.00 | $0.466 | $1.50 | $4.42 |
| Blended $/M | $0.166 | $0.875 | $3.75 | $0.262 | $0.70 | $2.16 |
The median new model costs $0.875/M blended against $0.70/M for the older catalog, about 25% more; the median output price is $2.50 against $1.50, and the output-to-input ratio is 4:1 in both groups. Restricting the baseline to the 264 older models that have snapshots in the window gives a median of $0.746, a 17% gap, so the direction does not depend on that choice.
The median hides the shape, though.
| Share of models | Q3 cohort (n = 91) | Older catalog (n = 338) |
|---|---|---|
| Blended price below $0.25 | 33.0% (30) | 23.7% (80) |
| Blended $0.25 – $1 | 17.6% (16) | 37.0% (125) |
| Blended $1 – $5 | 37.4% (34) | 28.1% (95) |
| Blended $5 and above | 12.1% (11) | 11.2% (38) |
| Context window 1M or more | 59.3% (54) | 18.3% (62) |
| Context window 200K or less | 12.1% (11) | 54.1% (183) |
| Median context window | 1,000,000 | 200,000 |
| Hugging Face ID listed (open-weight proxy) | 35.2% (32) | 46.2% (156) |
| Median output : input price ratio | 4.0 | 4.0 |
The new cohort is a barbell. A third of it (33%) costs under $0.25/M blended, against 24% before; the $0.25 to $1 middle band shrank from 37% to 18%; and the top quartile starts at $3.75 instead of $2.16. The share above $5 barely moved (12% against 11%). By month, the median blended price was $1.30 for July's 25 models, $0.87 for August's 23 and $0.88 for September's 43.
The structural change is context. The median new model offers 1,000,000 tokens against 200,000 for the older catalog, and 59% of new models offer 1M or more against 18%. Only 12% of new models stop at 200K, against 54%. In the older catalog a 1M window was the exception; in this cohort it is the default.
A second pattern: vendors held prices flat across generations and let the score rise. Google's Gemini 3.6, 3.7 and 3.8 Flash all list at $0.75/$3.75 while AA climbs from 34.3 to 39.4 to 41.2 (+20%). Meta's Muse Spark 1.1, 1.2 and 1.3 all list at $1.25/$4.25 while AA climbs from 34.3 to 39.8 to 45.2 (+32%). Anthropic did something different: Claude Opus 5 ($5/$25, AA 50.7) outscores the older Claude Fable 5 (49.7) at half its $10/$50 price.
5. Best new value picks
Raw AA per dollar rewards whatever is cheapest, so we use a stricter test. A model is on the frontier if no other scored model in the catalog has an equal or higher AA at an equal or lower blended price. First, the headline: among the 37 scored models with AA of 30 or more, 22 are Q3 newcomers, and in every band the newcomers are cheaper at the median.
| AA band | Q3 models (median blended) | Older models (median blended) |
|---|---|---|
| 45 and above | 5 ($10.00) | 1 ($20.00) |
| 40 – 45 | 8 ($3.00) | 2 ($10.00) |
| 30 – 40 | 9 ($1.50) | 12 ($4.66) |
| below 30 | 9 ($0.16) | 190 ($0.54) |
Here are all 31 scored newcomers, sorted by AA. "Beaten by another Q3 model" means a newer sibling or rival matches or exceeds its score for the same money or less; "beaten by an older model" means a pre-July model does. † marks the one model priced from its last recorded list price because it has no snapshot in the window.
| Model | AA | Input $/M | Output $/M | Blended $/M | Ctx | Position |
|---|---|---|---|---|---|---|
| Claude Fable 5.1 | 53.4 | $10.00 | $50.00 | $20.00 | 1M | On the frontier |
| GPT-6 Astra | 52.8 | $10.00 | $50.00 | $20.00 | 1M | Beaten by another Q3 model |
| Claude Opus 5 | 50.7 | $5.00 | $25.00 | $10.00 | 1M | On the frontier |
| GPT-5.6 Sol | 47.1 | $2.00 | $10.00 | $4.00 | 1M | On the frontier |
| Muse Spark 1.3 | 45.2 | $1.25 | $4.25 | $2.00 | 1M | On the frontier |
| GLM 5.3 | 44.9 | $1.40 | $4.40 | $2.15 | 1M | Beaten by another Q3 model |
| Grok 4.6 | 44.4 | $2.00 | $6.00 | $3.00 | 500K | Beaten by another Q3 model |
| Kimi K3 | 43.8 | $3.00 | $15.00 | $6.00 | 1M | Beaten by another Q3 model |
| GPT-5.6 Terra | 42.3 | $2.00 | $12.00 | $4.50 | 1M | Beaten by another Q3 model |
| GLM 5.3 Flash | 41.9 | $0.09 | $0.50 | $0.19 | 1M | On the frontier |
| Gemini 3.8 Flash | 41.2 | $0.75 | $3.75 | $1.50 | 1M | Beaten by another Q3 model |
| Qwen3.8 Max† | 40.3 | $2.00 | $6.00 | $3.00 | 1M | Beaten by another Q3 model |
| Qwen3.8 2.4T A95B | 40.0 | $2.00 | $6.00 | $3.00 | 1M | Beaten by another Q3 model |
| Muse Spark 1.2 | 39.8 | $1.25 | $4.25 | $2.00 | 1M | Beaten by another Q3 model |
| Gemini 3.7 Flash | 39.4 | $0.75 | $3.75 | $1.50 | 1M | Beaten by another Q3 model |
| Grok 4.5 | 39.1 | $2.00 | $6.00 | $3.00 | 500K | Beaten by another Q3 model |
| GPT-5.6 Luna | 37.5 | $0.20 | $1.20 | $0.45 | 1M | Beaten by another Q3 model |
| DeepSeek V4 Pro 0813 | 36.3 | $0.58 | $1.74 | $0.87 | 1M | Beaten by another Q3 model |
| DeepSeek V4 Flash 0731 | 34.5 | $0.04 | $0.28 | $0.10 | 1M | On the frontier |
| Gemini 3.6 Flash | 34.3 | $0.75 | $3.75 | $1.50 | 1M | Beaten by an older model |
| Muse Spark 1.1 | 34.3 | $1.25 | $4.25 | $2.00 | 1M | Beaten by an older model |
| Qwen3.8 27B | 33.9 | $0.42 | $3.00 | $1.06 | 1M | Beaten by an older model |
| Solar Pro 4 | 28.2 | $0.09 | $0.36 | $0.16 | 524K | Beaten by another Q3 model |
| Inkling Small | 26.1 | $0.45 | $1.20 | $0.64 | 524K | Beaten by an older model |
| Hy3 | 25.8 | $0.083 | $0.33 | $0.14 | 262K | Beaten by another Q3 model |
| Inkling | 25.5 | $1.00 | $4.05 | $1.76 | 524K | Beaten by an older model |
| Ling 3.0 Flash | 24.9 | $0.021 | $0.063 | $0.032 | 262K | On the frontier |
| Gemini 3.5 Flash Lite | 22.7 | $0.30 | $2.50 | $0.85 | 1M | Beaten by an older model |
| LongCat 2.0 | 19.7 | $0.30 | $1.20 | $0.52 | 1M | Beaten by an older model |
| Nemotron 3.5 Lightning | 13.6 | $0.08 | $0.20 | $0.11 | 262K | Beaten by an older model |
| Granite 4.2 8B | 12.4 | $0.06 | $0.25 | $0.11 | 131K | Beaten by an older model |
Seven of the 31 sit on the frontier, and they are seven of the eight frontier points in the entire catalog; the eighth is an older, tiny Ling 2.6 Flash at AA 9.7. The picks follow from that.
Best budget model with real capability: GLM 5.3 Flash. AA 41.9 at $0.09/$0.50 (blended $0.19, 1M context). It out-scores Gemini 3.8 Flash (41.2) at about one eighth of the price and out-scores older Claude Opus 4.7 (40.7, $10 blended) at roughly one fiftieth. Its daily blended price ranged from $0.07 to $0.24, and it stays on the frontier on its most expensive day.
Cheapest credible: DeepSeek V4 Flash 0731. AA 34.5 at $0.04/$0.28 (blended $0.10, 1M context). Its price is the most volatile of the group, from $0.05 to $0.33 blended depending on the day, so budget with the upper end. At the very bottom, Ling 3.0 Flash (AA 24.9, $0.021/$0.063) is the cheapest scored model.
Mid-price: Muse Spark 1.3. AA 45.2 at $1.25/$4.25 (blended $2.00), the highest score below $4 blended, narrowly ahead of GLM 5.3 (44.9 at $2.15).
Premium: GPT-5.6 Sol and Claude Opus 5. Sol scores 47.1 at $2/$10; Opus 5 scores 50.7 at $5/$25. Both sit on the frontier, and nothing cheaper reaches them.
OpenAI on a budget: GPT-5.6 Luna. AA 37.5 at $0.20/$1.20. It is not on the frontier (GLM 5.3 Flash beats it on both axes) but it is the best-scoring OpenAI model under $1. Its successor, GPT-6 Luna, lists at $0.10/$0.50 and has no score yet.
By budget. Below $0.25 blended, GLM 5.3 Flash is the only scored model above AA 40. Between $0.25 and $3, Muse Spark 1.3 leads. Between $3 and $10, GPT-5.6 Sol ($4) and Claude Opus 5 ($10) are the only frontier models. Above $10, only Fable 5.1 scores higher than Opus 5, by 5.3%.
6. Expensive newcomers and overpriced launches
The flagship premium is steep. Claude Fable 5.1 (AA 53.4) and GPT-6 Astra (52.8) both list at $10/$50, a $20 blended price. Against Opus 5 at $10 blended, that is +5.3% and +4.2% on AA for double the money. Astra is also marginally behind Fable 5.1 at an identical price.
Beaten by cheaper newer models. Kimi K3 (AA 43.8) has a median blended price of $6.00 and four newer models beat it, among them Muse Spark 1.3 (45.2 at $2.00) and GLM 5.3 (44.9 at $2.15); its cheapest day, at $2.77 blended, was still above both. GPT-5.6 Terra (42.3 at $4.50) is beaten by its own sibling Sol (47.1 at $4.00). Grok 4.5, Qwen3.8 Max and Qwen3.8 2.4T A95B all cost $3.00 blended for AA of 39 to 40, while GLM 5.3 Flash scores 41.9 at $0.19.
Beaten by older models. Inkling (25.5 at $1.76) is beaten by 12 older models, including MiniMax M3 (29.6 at $0.525). Gemini 3.5 Flash Lite (22.7 at $0.85) is beaten by DeepSeek V4 Flash 0423 (24.8 at $0.084), about ten times cheaper. LongCat 2.0 (19.7 at $0.52) is beaten by 12 older models. AA measures one thing, so these models may still win on tool use, language coverage or latency, but the burden of proof is on them.
Unscored and expensive. Several launches, most of them late in the quarter, carry premium prices and no AA score yet: Claude Opus 5 Fast at $10/$50, Fugu Ultra v2 at $5/$30, Claude Opus 5.5 at $4/$20, Qwen3.8 Max Prime at $4/$12 and Ember-1 at $3/$15. Speed tiers carry the starkest markups: Opus 5 Fast is priced at twice standard Opus 5, and MiMo-V2.6-Pro-UltraSpeed at $4.35/$8.70 is exactly ten times MiMo-V2.6-Pro at $0.435/$0.87.
7. What to watch
Scores are still arriving. Thirty-two of the 91 models were first seen after the 8 September AA snapshot, among them GPT-6 Luna ($0.10/$0.50), GPT-6.1 Sol ($2/$10), Claude Sonnet 5.5 ($2/$10), Grok 4.7 ($2/$6) and MiMo-V2.6-Pro. Several could land on the frontier or knock current members off it, so recheck the best-value ranking once scores appear.
Batch mirrors will keep inflating counts. Expect more bulk batch listings, and treat any "new models this month" figure that does not de-duplicate them with suspicion.
Endpoint-driven price swings. Kimi K3, DeepSeek V4 Flash 0731, GLM 5.3 and GLM 5.3 Flash all moved by a factor of two or more across the 31-day window. Check the live page on the models list before committing a budget to a model's cheapest day.
Using this list. Treat the frontier as a shortlist, not a verdict: AA is one general index, so rerun your own evaluation on two or three candidates and take the cheapest that passes. Budget with the upper end of a model's daily price range rather than its median; for DeepSeek V4 Flash 0731 that range ($0.05 to $0.33 blended) spans more than a factor of six.
The short version: the quarter produced 95 distinct models rather than 224, with September the real peak. The price distribution widened at both ends, one million tokens of context became the norm, and seven of the eight points on the price-versus-intelligence frontier now belong to models first seen in Q3.