The short answer for most teams: use Claude Opus 5.5 for serious coding and agent work, and test GPT-6.1 Sol as a cheaper option. Use Gemini 3.8 Flash for fast, cheap work that mixes text, images, audio and video. Use DeepSeek-V4.1-Flash or GPT-6 Luna for bulk jobs. Use Gemma 4, Qwen3.8-27B or Sarvam when you need to run the model yourself or serve Indian languages. Each section below gives the evidence for its picks and names one model to avoid.
First principles
Every pick below follows from a few facts about what a model call is made of.
- You pay per token, but you budget per task. A task’s cost is the price per token times the tokens the task uses, and reasoning models spend output tokens thinking before they answer. A model with a higher price per token can still cost less per task if it uses fewer tokens, and the reverse. That is why we quote cost per task wherever an evaluation published it.
- A model can only use what fits in its window and what it can read. The context window caps how much one request can hold, and the accepted input types cap what the model can see. A report that does not fit has to be split into chunks, and a text-only model cannot read a chart without a conversion step.
- Weights must fit in memory to run. At 16-bit precision each parameter takes two bytes, so a 31B model needs about 62GB for its weights alone, and a model with trillions of parameters needs several machines. “Open weights” and “runs on your hardware” are different claims.
- Where the model runs decides where your data goes. A hosted application programming interface (API) sends every prompt to the vendor’s servers, under the vendor’s terms and the law of the country they sit in. Self-hosting open weights keeps the data with you, at the cost of running the hardware.
- A benchmark measures its own tasks, not yours. A leaderboard score is evidence about the jobs it tests. Where the test differs from the job, as when a general knowledge score is used to pick a long-document model, we say so and treat the pick as a reasoned bet.
Coding agents
Pick: Claude Opus 5.5. It ranks first on the Arena web development board (1820, updated 29 September 2026) and first on the Artificial Analysis Intelligence Index (58). It is expensive per task, so budget for it.
Consider: GPT-6.1 Sol. At max effort it scored 52 on the same index at $0.72 per task, against 58 at $5.98 for Opus 5.5 at max effort. OpenAI describes it as built for complex coding, but we found no independent coding result for it. It was one day old when we reviewed it, so check it on your own tasks.
Avoid: Gemini 3.1 Pro. It is still a preview model, and it scored 33.44% on the Vals Index against about 67% for the Claude 5.5 models.
Long documents and research
Pick: GPT-6 Astra. It leads Scale’s Humanity’s Last Exam board (54.80) and Epoch AI’s Capabilities Index (166.6), with a 1,050,000 token window. Prompts over 272K tokens cost more. Those two scores measure general knowledge and capability, not long-document work, and we found no independent long-context test of it, so treat the pick as a reasoned bet.
Pick: Gemini 3.8 Flash. It reads up to 1,048,576 tokens, including Portable Document Format (PDF) files, audio and video, at $0.75 input per million tokens until the end of 2026.
Claude Opus 5.5 is also a candidate here, with a 1M token window, but Artificial Analysis notes it still trails other models on its long-context test, Artificial Analysis Long Context Reasoning (AA-LCR).
Avoid: Sarvam 30B. Its 65,536 token window is too short for long reports or codebases without chunking.
High volume at low cost
Pick: DeepSeek-V4.1-Flash. It scored 51.32% on the Vals Index at $0.33 per test. Off-peak prices are $0.15 input and $0.60 output per million tokens. Its hosted API stores personal data in China, so self-host the MIT-licensed weights if that matters.
Pick: GPT-6 Luna. It scored 51.22% on Vals at $0.43 per test and costs $0.10 input and $0.50 output per million tokens.
Avoid: Kimi K3. Its API charges $15 per million output tokens, and Vals shows 50.70% at $6.97 per test, a lower score than the two picks at many times the cost.
Structured extraction and classification
Pick: Gemini 3.8 Flash. It supports structured outputs, batch and caching, and reads PDFs and images directly, which removes a conversion step.
Pick: IBM Granite 4.2 8B (self-hosted). Reasoning can be switched off per request, which keeps simple labelling fast. It is Apache 2.0 and runs on modest hardware. Its scores are vendor-reported only.
Avoid: Claude Opus 5.5. Its thinking cannot be turned off and forcing a specific tool call returns an error, which removes a common way to force a schema. We did not test how this affects cost or accuracy.
Multilingual and Indic languages
Pick: Sarvam 30B or 105B. Sarvam built them to cover all 22 scheduled Indian languages across 12 scripts, under Apache 2.0. There is no independent evaluation yet, so test in your target languages.
Pick: Gemma 4. Its model card lists more than 140 languages, and it is a practical base for fine-tuning.
Avoid: MiMo-V2.6. Its model card lists only English and Chinese.
Images and video input
Pick: Gemini 3.8 Flash. It accepts images, video, audio and PDF in one request, and is fast.
Pick: Qwen3.8-Max. It ranks second on the Arena vision board (1301, updated 28 September 2026).
Avoid: gpt-oss-120b. It is a text-only model.
Voice and real time
Pick: Gemini 3.8 Live. It is Google’s default low-latency voice model, with audio input at $3.00 and audio output at $12.00 per million tokens.
Pick: OpenAI gpt-live-1 or gpt-realtime-2.1. gpt-live-1 is a full-duplex voice model at $0.05 per minute, with model and tool use charged separately. gpt-realtime-2.1 charges $32 per million audio input tokens and $64 per million audio output tokens.
We found no independent evaluation of voice models for this edition, so these picks rest on vendor documentation.
Avoid: Claude Opus 5.5. Anthropic lists text and image input only, with no audio.
Running on your own hardware
Pick: Gemma 4. Sizes run from phones (E2B and E4B) through a 16GB laptop (12B) to a single 80GB graphics processing unit (GPU) for the 31B, all under Apache 2.0.
Pick: Qwen3.8-27B. It is a 27B dense Apache 2.0 model that reads images, and Artificial Analysis gives it 34 on its open-weights page. For comparison, its main leaderboard gives Gemma 4 31B 19.
Avoid: Kimi K3. Its weights are open, but its model card lists 2.8 trillion (2.8T) total parameters, far more than any model here that runs on one GPU.
Reasoning heavy tasks
Pick: GPT-6 Astra. Epoch AI records 94% on FrontierMath Tiers 1 to 3 (v2) and ranks it first on its Capabilities Index.
Pick: Claude Opus 5.5. Epoch AI records 91% on FrontierMath Tiers 1 to 3 (v2), and Artificial Analysis measured 61.4% on Humanity’s Last Exam.
Avoid: Mistral Medium 3.5. Artificial Analysis gives it 14 on its Intelligence Index and Epoch AI 141.4 on its Capabilities Index, far below the picks.
How we reviewed
These reviews are desk research. We read vendor documentation, model cards, pricing pages and announcements, and named independent evaluations: Artificial Analysis, Arena (formerly LMArena), the Vals Index, the Epoch AI Benchmarking Hub, ARC Prize, Scale’s Humanity’s Last Exam board and Model Evaluation and Threat Research (METR). We read all of them on 30 September 2026, and each review lists its sources. Numbers marked vendor-reported come from the vendor and were not checked by anyone else.
Velofy did not run its own tests of these models for this edition. Several evaluations we would have liked to use were out of date: the public SWE-bench and Aider leaderboards and METR’s time-horizon page had no results for the newest models when we read them.
Prices, versions and rankings change quickly. Some models reviewed here were only days old when we read these sources. Check the vendor’s pricing page before you commit, and run your own evaluation on your own task.
Sources
Read on 30 Sept 2026.
- LLM Leaderboard (Artificial Analysis)
- Claude Opus 5.5 takes the top spot on the Artificial Analysis Intelligence Index
- Top Open Source AI Models (Artificial Analysis)
- WebDev AI Leaderboard (Arena)
- Vision leaderboard (Arena)
- Vals Index
- Epoch AI Benchmarking Hub
- Humanity's Last Exam leaderboard (Scale Labs)
- Claude Opus 5.5
- Models overview (Claude)
- GPT-6.1 Sol model page
- GPT-6 Astra model page
- GPT-6 Luna model page
- gpt-realtime-2.1 model page
- gpt-live-1 model page
- OpenAI API pricing
- openai/gpt-oss-120b · Hugging Face
- Gemini 3.8 Flash (Gemini API docs)
- Gemini 3.8 Live
- Gemini Developer API pricing
- Gemini 3.1 Pro preview
- Gemma 4 model card
- Gemma 4: Byte for byte, the most capable open models
- Models & Pricing | DeepSeek API Docs
- Qwen/Qwen3.8-27B · Hugging Face
- Kimi-K3 · Hugging Face
- Model Inference Pricing Explanation (Kimi API Platform)
- ibm-granite/granite-4.2-8b · Hugging Face
- Open-Sourcing Sarvam 30B and 105B
- sarvamai/sarvam-30b · Hugging Face
- XiaomiMiMo/MiMo-V2.6-Flash-RL · Hugging Face
- Mistral Medium 3.5 - Mistral AI
- GPT-6 Astra (Epoch AI)
- Claude Opus 5.5 (Epoch AI)
- DeepSeek Privacy Policy
- DeepSeek-V4.1-Flash · Hugging Face
- OpenAI launches GPT-6.1 Sol (TechCrunch)
- ARC Prize results
- METR time horizons
- SWE-bench leaderboards
- Aider LLM leaderboards