Agnes 2.5 Pro Theta vs Inkling
Full profiles: Agnes 2.5 Pro Theta (Sapiens AI) and Inkling (Thinking Machines): every benchmark, price and rank.
Updated · Data from 11 sources
As of 21 Nov 2026, Agnes 2.5 Pro Theta and Inkling are neck and neck on quality (50.7 vs 50.2 on the BenchLeader Index, inside the ±10.5 margin), and Agnes 2.5 Pro Theta is 17× cheaper at $0.150 vs $2.57 per 1M tokens blended. Inkling leads in every category we measure both in. Agnes 2.5 Pro Theta streams ~1.5× faster.
Key facts (as of 21 Nov 2026)
- Agnes 2.5 Pro Theta scores 50.7 on the BenchLeader Index and Inkling scores 50.2 (BenchLeader, 21 Nov 2026).
- Agnes 2.5 Pro Theta costs $0.150 per 1M tokens blended (list price, 21 Nov 2026).
- Inkling costs $2.57 per 1M tokens blended (list price, 21 Nov 2026).
- Agnes 2.5 Pro Theta streams ~1.5× faster than Inkling (Artificial Analysis, 7-day median, 21 Nov 2026).
Which to pick for…
The pick's numbers come first. Category scores on the BenchLeader scale (50 is the average model, 15 points one standard deviation). High confidence means a lead of half a standard deviation or more; under one point is called either.
| Use case | Pick | Why | Confidence |
|---|---|---|---|
| Coding | Either | 45 vs 46 category score | Either |
| Reasoning | Inkling | 52 vs 43 category score | High |
| Knowledge | Inkling | 60 vs 40 category score | High |
| Chat quality (human preference) | – | Only Inkling has an Arena rating (59 category score) | Not comparable |
| Lowest cost | Agnes 2.5 Pro Theta | $0.150 vs $1.76 per 1M blended; agentic coding session $0.007 vs $0.076 | – |
| Fastest first answer | Inkling | 0.2 s (Inkling) vs 8.8 s (Agnes 2.5 Pro Theta) to the first answer token | – |
| High volume on a budget | Neither flagship | A smaller model is usually the better fit: Cheapest good LLMs, Fastest LLMs | – |
Key differences
- Agnes 2.5 Pro Theta and Inkling are neck and neck on quality (50.7 vs 50.2, within the ±10.5 margin).
- Inkling is stronger in coding, knowledge, reasoning.
- Agnes 2.5 Pro Theta is 17× cheaper ($0.150 vs $2.57 per 1M blended).
- Agnes 2.5 Pro Theta streams 1.7× faster (189 vs 110 tokens per second).
At a glance
| Spec | Agnes 2.5 Pro Theta | Inkling |
|---|---|---|
| Developer | Sapiens AI | Thinking Machines |
| Released | 7 Nov 2026 | 15 Jul 2026 |
| Context window (tokens) | 1,000,000 | 65,536 |
| Compared at | Agnes 2.5 Pro Theta | Inkling |
| Open weights | No | Yes |
| Price in / out / cached input per 1M | $0.100 / $0.300 / $0.010 | $1.00 / $4.05 / $0.170 |
Category by category
Category scores on the BenchLeader scale: 50 is the average evaluated model, 15 points is one standard deviation. A gap under one point is called level.
| Category | Agnes 2.5 Pro Theta | Inkling | Leader |
|---|---|---|---|
| Reasoning | 43.2 | 51.5 | Inkling |
| Coding | 45.3 | 46.2 | level |
| Agents & tools | – | 30.2 | not comparable |
| Maths | – | 54.1 | not comparable |
| Knowledge | 39.5 | 60.3 | Inkling |
| Human preference | – | 58.8 | not comparable |
| Multimodal | 60.4 | – | not comparable |
| Long context | 67.1 | – | not comparable |
On one chart
Each domain on the index's own scale, where 50 is an average model. Click a domain to see the benchmarks behind it. Pick up to 5 models from the list.
Not drawn, measured for fewer than two of these models: Agents & tools, Maths, Instruction following, Long context, Multimodal, Human preference.
Where sources disagree
Across the 0 benchmarks both have, Agnes 2.5 Pro Theta leads 0, Inkling leads 0. Every overall index that rates both puts Agnes 2.5 Pro Theta ahead.
| Overall index | Agnes 2.5 Pro Theta | Inkling | Leader |
|---|---|---|---|
| BenchLeader Index | 50.7 | 50.2 | neck and neck |
| AA Intelligence Index v5.0.1 | 24.0 | 11.0† | Agnes 2.5 Pro Theta |
Each publisher runs its own tasks, harnesses and effort settings, so their overall verdicts can differ; BenchLeader reconciles them on one scale. † measured at another effort setting of the same model (the publisher did not test the one compared here).
Price, speed and latency
List prices blended at three input tokens per output token. Per-task prices use the token counts on the methodology page. Cost, tokens and time per benchmark task are measured: Artificial Analysis’ averages over its whole Intelligence Index run.
| Measure | Agnes 2.5 Pro Theta | Inkling |
|---|---|---|
| Blended price per 1M tokens | $0.150 | $1.76 |
| Input / output per 1M | $0.100 / $0.300 | $1.00 / $4.05 |
| Cached input per 1M | $0.010 | $0.170 |
| Price per task: chat reply | $0.0001 | $0.0016 |
| Price per task: summarize a 30-page report | $0.0014 | $0.014 |
| Price per task: code edit | $0.0011 | $0.012 |
| Price per task: agentic coding session | $0.0072 | $0.076 |
| Price per task: structured extraction | $0.0003 | $0.0028 |
| Agentic coding session, input mostly cached | $0.0032 | $0.039 |
| Cost per benchmark task (AA Index) | $0.575 | – |
| Output tokens per task (AA Index) | 278k | – |
| Of which reasoning | 229k | – |
| Time per task (AA Index) | 24 min | – |
| Output speed | 189 tok/s | 110 tok/s |
| Time to first answer | 8.8 s | 0.2 s |
| Time to a complete answer, last measured 13 Nov 2026 | 16.9 s | – |
Each model is compared at its best-ranked reasoning-effort setting. Agnes 2.5 Pro Theta scores 50.7 and takes 9 s to its first answer token. Inkling scores 50.2 and takes 0 s to its first answer token; Inkling (xhigh) scores 49.9 in 18 s, Inkling (high) scores 44.0 in 10 s. Higher effort means more thinking before the answer: better scores, longer waits and more tokens billed. How reasoning effort changes scores and cost.
Every benchmark
134 benchmarks, grouped by category; each name links to the board and its leaders. † measured at another effort setting of the same model, shown for reference and not counted as a win.
| Metric | Publisher | Agnes 2.5 Pro Theta | Inkling |
|---|---|---|---|
| BenchLeader Index | BenchLeader | 50.7 | 50.2 |
| Coding score | BenchLeader | 45.3 | 46.2 |
| Knowledge score | BenchLeader | 39.5 | 60.3 |
| Long context score | BenchLeader | 67.1 | – |
| Multimodal score | BenchLeader | 60.4 | – |
| Reasoning score | BenchLeader | 43.2 | 51.5 |
| Agents & tools score | BenchLeader | – | 30.2 |
| Human preference score | BenchLeader | – | 58.8 |
| Maths score | BenchLeader | – | 54.1 |
| Blended price $/M | BenchLeader | $0.150 | $1.76 |
| Output speed | BenchLeader | 189 tok/s | 110 tok/s |
| Time to first answer | BenchLeader | 8.8 s | 0.2 s |
| Context window | BenchLeader | 1,000,000 tokens | 65,536 tokens |
| Reasoning | |||
| GPQA Diamond | Epoch AI Benchmarking Hub | – | 88.3%† |
| SimpleBench | SimpleBench | – | 50.0% |
| Arena Hard Prompts | Arena (LMArena) | – | 1463 |
| LiveBench 2.0 Reasoning | LiveBench | – | 78.3%† |
| LiveBench 1.0 Reasoning | LiveBench | – | 78.3%† |
| GPQA Diamond (AA) | Artificial Analysis | – | 87.2%† |
| Humanity's Last Exam (AA) | Artificial Analysis | 37.2% | 31.9%† |
| HLE-Diamond (AA) | Artificial Analysis | 16.8% | 10.6%† |
| GPQA Diamond (Vals) | Vals AI | – | 87.1% |
| ARC-AGI-1 | ARC Prize | – | 79.5% |
| ARC-AGI-2 | ARC Prize | – | 36.5% |
| CritPt | Artificial Analysis | 15.1% | 5.4%† |
| CritPt v2 hard subset (AA) | Artificial Analysis | 26.0% | 18.0%† |
| HLE-Diamond (Scale) | Scale AI SEAL | – | 11.2%† |
| Arena Expert | Arena (LMArena) | – | 1474 |
| Chess Puzzles | Epoch AI Benchmarking Hub | – | 21.0%† |
| LMCA | Epoch AI Benchmarking Hub | – | 37.6%† |
| DTBench | Epoch AI Benchmarking Hub | – | 87.5%† |
| Coding | |||
| SciCode | SciCode | – | 47.0%† |
| WeirdML | WeirdML | – | 32.3%† |
| FrontierCode | Cognition | – | 14.0% |
| Arena Coding | Arena (LMArena) | – | 1494 |
| Arena WebDev | Arena (LMArena) | – | 1414 |
| LiveBench 2.0 Coding | LiveBench | – | 48.7%† |
| LiveBench 1.0 Coding | LiveBench | – | 71.0%† |
Show all 134 benchmarks
| Metric | Publisher | Agnes 2.5 Pro Theta | Inkling |
|---|---|---|---|
| Coding | |||
| SciCode (AA) | Artificial Analysis | 56.0% | 47.0%† |
| LiveCodeBench | Vals AI | – | 85.5% |
| IOI | Vals AI | – | 14.9% |
| SWE-bench (Vals) | Vals AI | – | 77.6% |
| Code Migration | Vals AI | – | 11.8% |
| Code Migration v1.1 | Vals AI | – | 8.8% |
| ProgramBench | Vals AI | – | 0.0% |
| Vibe Code Bench 1-100 | Vals AI | – | 7.3% |
| Vibe Code Bench v1.1 | Vals AI | – | 19.2% |
| SWE-Bench Pro V2 | Scale AI SEAL | – | 89.9%† |
| SWE-Bench Pro V2 HARD | Scale AI SEAL | – | 56.9%† |
| ALE-Bench | Epoch AI Benchmarking Hub | – | 946 |
| FrontierSWE | Epoch AI Benchmarking Hub | – | 4.1%† |
| WeirdML v3 | WeirdML | – | 3.9%† |
| Terminal-Bench 4.0 (AA) | Artificial Analysis | 10.1% | 1.0%† |
| Terminal-Bench 2.1 (AA) | Artificial Analysis | – | 55.1%† |
| GameBench 2 Overall | DuelLab GameBench 2 | – | 25.6 |
| GameBench 2 by reasoning setting | DuelLab GameBench 2 | – | 26.8† |
| Design Arena Code | Design Arena | – | 1186 |
| FrontierCode 1.1 (AA) | Artificial Analysis | – | 15.8%† |
| FrontierCode 2.0 (AA) | Artificial Analysis | 16.6% | 3.9%† |
| Agents & tools | |||
| Terminal-Bench | Terminal-Bench | – | 0.0%† |
| APEX-Agents | Mercor | – | 33.8% |
| Arena Agent | Arena (LMArena) | – | -18.5 |
| LiveBench Long Agentic | LiveBench | – | 9.8%† |
| LiveBench Browser | LiveBench | – | 12.0%† |
| LiveBench Agentic Coding | LiveBench | – | 49.4%† |
| Terminal-Bench 2.1 (Vals) | Vals AI | – | 47.6% |
| MCP Atlas | Scale AI SEAL | – | 76.0%† |
| τ³-bench Banking | τ²-bench | – | 25.0%† |
| GDPval-AA v2.1 | Artificial Analysis | 1507 | 1080† |
| τ³-Banking (AA) | Artificial Analysis | – | 29.1%† |
| Analyst Agent (AA) | Artificial Analysis | – | 23.8%† |
| CyberBench | Vals AI | – | 66.8% |
| Finance Agent v2 | Vals AI | – | 46.6% |
| Finance Agent v2.1 | Vals AI | – | 49.9% |
| Harvey's Legal Agent Benchmark v1.1 | Vals AI | – | 3.8% |
| Harvey's Legal Agent Benchmark v1 | Vals AI | – | 2.1% |
| Legal Research Bench | Vals AI | – | 28.4% |
| SkillsBench | Vals AI | – | 26.1% |
| SREBench | Vals AI | – | 0.0% |
| Tax Agent Bench | Vals AI | – | 15.3% |
| Terminal-Bench 4.0 (Vals) | Vals AI | – | 0.5% |
| Arena Agent Task Completion | Arena (LMArena) | – | -26.9 |
| AutomationBench | Artificial Analysis | 48.8% | 5.0%† |
| GDP.pdf (AA) | Artificial Analysis | 14.2% | 12.8%† |
| Harvey LAB-AA v1.1 | Harvey | – | 0.0%† |
| EnterpriseOps-Gym | Artificial Analysis | – | 38.0%† |
| AA-Briefcase v1.1 | Artificial Analysis | 1496 | 833† |
| GDPx | GDPx by SemiAnalysis | – | 2.3%† |
| Design Arena Agentic | Design Arena | – | 1162 |
| Terminal-Bench-Science 0.1 (AA) | Artificial Analysis | 3.8% | 0.0%† |
| Maths | |||
| FrontierMath Tiers 1–3 | Epoch AI Benchmarking Hub | – | 33.3%† |
| FrontierMath Tier 4 | Epoch AI Benchmarking Hub | – | 4.9%† |
| OTIS Mock AIME | Epoch AI Benchmarking Hub | – | 88.9%† |
| ProofBench | Vals AI | – | 0.0% |
| LiveBench Mathematics | LiveBench | – | 88.4%† |
| Arena Maths | Arena (LMArena) | – | 1479 |
| Knowledge | |||
| SimpleQA Verified | Epoch AI Benchmarking Hub | – | 40.3%† |
| LiveBench 2.0 Data Analysis | LiveBench | – | 72.8%† |
| LiveBench Research | LiveBench | – | 60.3%† |
| LiveBench 1.0 Data Analysis | LiveBench | – | 72.8%† |
| AA-Omniscience v1.1 | Artificial Analysis | -11.2 | -1.5† |
| MMLU-Pro | Vals AI | – | 86.3% |
| LegalBench | Vals AI | – | 83.0% |
| CorpFin | Vals AI | – | 68.6% |
| TaxEval | Vals AI | – | 75.3% |
| Excel Modeling Benchmark | Vals AI | – | 38.4% |
| Excel Modeling Benchmark v1.1 | Vals AI | – | 41.0% |
| MedCode | Vals AI | – | 41.2% |
| MedScribe | Vals AI | – | 85.4% |
| Public Benefits Bench v1.1 | Vals AI | – | 58.5% |
| Arena Text Factuality | Arena (LMArena) | – | 1452 |
| AA-Omniscience v1.1: accuracy | Artificial Analysis | 21.3% | 39.3%† |
| AA-Omniscience v1.1: non-hallucination | Artificial Analysis | 58.6% | 32.8%† |
| Instruction following | |||
| LiveBench Language | LiveBench | – | 73.5%† |
| LiveBench Instruction Following | LiveBench | – | 70.1%† |
| Arena Instruction Following | Arena (LMArena) | – | 1427 |
| Human preference | |||
| Arena Text | Arena (LMArena) | – | 1443 |
| EQ-Bench 4 | EQ-Bench | – | 1212 |
| EQ-Bench Creative Writing v3 | EQ-Bench | – | 1609 |
| EQ-Bench Longform Writing | EQ-Bench | – | 72.5 |
| Arena Creative Writing | Arena (LMArena) | – | 1389 |
| Arena Multi-turn | Arena (LMArena) | – | 1445 |
| Arena Longer Queries | Arena (LMArena) | – | 1442 |
| Multimodal | |||
| LiveBench Design | LiveBench | – | 55.2%† |
| MMMU-Pro | Artificial Analysis | 78.4% | 73.5%† |
| MMMU-Pro (Vals) | Vals AI | – | 78.5% |
| MortgageTax | Vals AI | – | 64.0% |
| AudioMC | Scale AI SEAL | – | 56.6%† |
| AudioMC (text output) | Scale AI SEAL | – | 56.6%† |
| Code2Video-AA | Artificial Analysis | – | 467† |
| AA-3D | Artificial Analysis | – | 591† |
| Long context | |||
| AA-LCR v1.1 | Artificial Analysis | 84.7% | 77.3%† |
| MLCR | Artificial Analysis | – | 6.1%† |
| Composite | |||
| Epoch Capabilities Index | Epoch AI Benchmarking Hub | – | 148.4 |
| LiveBench 2.0 | LiveBench | – | 44.3%† |
| LiveBench 1.0 | LiveBench | – | 71.9%† |
| AA Intelligence Index v5.0.1 | Artificial Analysis | 24.0 | 12.5† |
| AA Intelligence Index v4.3.2 | Artificial Analysis | – | 25.0† |
| AA Progress Index v1 | Artificial Analysis | 304.2 | 281.4† |
| Vals Index | Vals AI | – | 29.0 |
| Vals Finance Index | Vals AI | – | 38.3% |
| Vals Multimodal Index | Vals AI | – | 49.4% |
| Vals Router Index | Vals AI | – | 29.0% |
| Vals Legal Index | Vals AI | – | 54.2% |
| Safety & honesty | |||
| FORTRESS | Scale AI SEAL | – | 25.0%† |
| DistressBench | Scale AI SEAL | – | 78.5% |
| Metadata | |||
| AA Openness Index | Artificial Analysis | – | 38.9† |
Data as of 2026-11-21. Best configuration of each model; every score links to its source on the model pages.
DuelLab GameBench 2, release gb2-20261119-1d7fd529faf6 (https://benchmarks.duellab.org/releases/gb2-20261119-1d7fd529faf6/), CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). Scores mapped to BenchLeader model names; not endorsed by DuelLab.
What changed
- Agnes 2.5 Pro Theta: 1 new result (AA Progress Index v1).
- Inkling: 1 score revision.
- Inkling: 2 score revisions.
- Inkling: BenchLeader Index 47.8 → 50.4.
- Inkling: 1 score revision.
- Agnes 2.5 Pro Theta: first results.
- Inkling: 5 score revisions.
- Inkling: 1 score revision.
- Inkling: 1 score revision.
Related comparisons
- Agnes 2.5 Pro Theta vs Nemotron 3 Super 120B A12B
- Agnes 2.5 Pro Theta vs GLM 4.7 Flash
- Agnes 2.5 Pro Theta vs Ling 3.1 Flash
- Inkling vs Inkling 2 Mini
- Inkling vs Mistral Large 4.0 OSS
- Agnes 2.5 Pro Theta vs Qwen3.5-9B
- Agnes 2.5 Pro Theta vs Apodex 1.1
Rankings: Best LLM overall · Best LLM for agents · Best cheap LLMs
Agnes 2.5 Pro Theta vs Inkling: questions
- Is Agnes 2.5 Pro Theta better than Inkling?
- Agnes 2.5 Pro Theta and Inkling are neck and neck on quality (50.7 vs 50.2, within the ±10.5 margin). The BenchLeader Index combines every independent quality benchmark; the two are within the margin of error as of 21 Nov 2026, so check the category scores for your use.
- Agnes 2.5 Pro Theta vs Inkling: which is cheaper, and will that change?
- Agnes 2.5 Pro Theta is cheaper today: $0.150 against $2.57 per million tokens, blended at three input tokens per output token. No price change has been announced for either.
- Is Agnes 2.5 Pro Theta or Inkling better for coding?
- They are level in coding (45 vs 46 on the category index).
- Which is faster, Agnes 2.5 Pro Theta or Inkling?
- Agnes 2.5 Pro Theta streams ~1.5× faster (189 vs 110 output tokens per second, Artificial Analysis 7-day median).
- Which is better for long documents?
- Agnes 2.5 Pro Theta accepts more context: 1,000,000 against 65,536 tokens for Inkling.
More questions (2)
- Which has lower latency?
- Inkling answers first: 0.2 s against 8.8 s to the first answer token, as Artificial Analysis measures it. Reasoning effort drives most of this: a higher setting thinks longer before it answers.
- When was Agnes 2.5 Pro Theta released?
- Agnes 2.5 Pro Theta was released on 7 Nov 2026.