AI models for programming and code

From snippets to repositories

A solution producing correct code in a small example may lose context across files or add unnecessary dependencies. Evaluate the change within the project, using its language version, environment and conventions. Preserve the starting state and review the diff. A convincing explanation does not replace compilation, tests or inspection of side effects.

A concrete example

To fix an endpoint, prepare a request reproducing the failure and a check of expected behavior. Give the model the necessary context, apply the patch in an isolated environment and compare the result. Check invalid inputs and permissions as well as the success case.

What to check

Compare time to a validated solution, cost, attempts and patch maintenance. Code benchmarks depend on the harness, tools and budget; do not conflate a single answer with an agent running multiple executions.

Models in this category

29 models with profiles. Classification describes capabilities, not a quality ranking.

OpenAI

GPT‑6.1 Terra

Proprietary OpenAI model released on October 26, 2026 for interpreting images, using applications and writing text. It accepts text and images and returns text; generating images through a tool is not a native output modality. It is available through the API as gpt-6.1-terra and in ChatGPT Work and Codex depending on the plan. It is not released as a downloadable checkpoint.

Text and reasoning · Paid proprietary NVIDIA

Nemotron 3.5 Bolt E4B

Compact NVIDIA model released on October 23, 2026, post-trained from Gemma-4-E4B-it. It has approximately 8 billion total and 4 billion effective parameters; E4B does not mean an MoE with four billion active parameters. We group BF16 and GGUF Q4_K_M in one profile, with results separated by precision. The documented modality is text to text: targeting voice assistants does not make it an audio recognition or synthesis system.

Text and reasoning · API and open weights Recursive

Recursive Flint

Recursive model released on October 21, 2026 to investigate repositories, fix bugs, develop features and write tests with local agents. It is a mixture-of-experts Transformer with 39.8 billion total parameters and approximately 2.5 billion active per token, including the output head. It accepts and generates text; it does not process images or audio. This profile groups the BF16 checkpoint and official INT4 and NVFP4 distributions without counting them as independent models.

Code · API and open weights NVIDIA

Nemotron Labs Dory 1.0 4B

NVIDIA text model released on October 16, 2026, with 4 billion parameters and a recurrent architecture. It reuses a block of layers to adjust compute per token without switching checkpoints. English and a validated 256K-token context are documented. It does not generate images, audio or video.

Text and reasoning · Open weights Meta

Muse Spark Open

Meta multimodal model released on October 15, 2026, with 816 billion total and 42 billion active parameters. It accepts text, images, audio and video and returns text. It has a combined one-million-token context and BF16 and NVFP4 variants.

Multimodal · Restricted weights Google

Gemini 4 Argon

Google model generally available since October 13, 2026. It accepts text, images, audio and video and generates text, including code; it does not generate images, audio or video. It uses a MoE architecture and a shared 1,048,576-token window. Robotics applications use this same model, not a separate endpoint.

Text and reasoning · Paid proprietary Microsoft

AesCode 32B

Microsoft's multimodal family turns content requirements and optional visual references into editable HTML/CSS documents. The 8B and 32B checkpoints derive from Qwen3-VL-8B-Instruct and Qwen3-VL-32B-Instruct respectively. They produce code, not raster images or video files. The reference provides composition and style, while the prompt remains the source of truth for text, figures and relationships. The verified public release is October 11, 2026; earlier repository creation is not its release date.

Code · Open weights Microsoft

AesCode 8B

Code · Open weights Anthropic

Claude Haiku 5.5

Anthropic model released on October 7, 2026 for frequent, cost- and latency-sensitive tasks. It accepts text and images and returns text, including code; it does not generate images, audio or video. Its one-million-token window and adaptive thinking expand the Haiku family without making it a universal replacement for Sonnet or Opus on complex work.

Text and reasoning · Paid proprietary Anthropic

Claude Sonnet 5.5

Sonnet update for everyday tasks, code, and documents, with improved speed and a million tokens of context.

Text and reasoning · Paid proprietary Aleph Alpha

Kolibri 1

Aleph Alpha's text and code model, officially released on October 3, 2026, focused on German and English. Its MoE architecture has 78.1 billion total parameters and activates about 3.46 billion per token. It supports configurable reasoning and tool calling. FP8 and BF16 weights are released under Apache 2.0; they are two precisions of the same family, not two independent models.

Text and reasoning · Open weights OpenAI

GPT‑6.1 Sol

OpenAI model for complex coding, computer use and professional work with text and image inputs. The provider describes performance close to Astra at a lower cost; assess that trade-off on your own tasks.

Text and reasoning · Paid proprietary Anthropic

Claude Opus 5.5

Next-generation Opus for agentic coding, computer use, and extended professional work, with a one-million-token context.

Text and reasoning · Paid proprietary Z.AI

GLM‑5.3

Z.AI open model focused on software engineering, long-running agents, and technical reasoning.

Code · API and open weights Z.AI

GLM‑5.3‑Flash

Multimodal variant with 320.000 billion total parameters and 18.000 billion active parameters, focused on efficiency.

Multimodal · API and open weights xAI

Grok 4.7

xAI model for coding, agents, and knowledge work with text and image input.

Text and reasoning · Paid proprietary Alibaba Qwen

Qwen3.8-Max

Qwen3.8 flagship family for software engineering, agents, and visual understanding with a one-million-token context window.

Text and reasoning · Paid proprietary Anthropic

Claude Opus 4.5

Previous Opus version focused on code, agents, and computer use, retained for migration and historical comparison.

Text and reasoning · Paid proprietary

Pagination Page 1 of 2 Showing 1–18 of 29

ANALYSIS

Related analysis

GUIA

Retries in AI APIs: How to Recover from Failures Without Duplicating Requests or Costs

A safe retry policy does not repeat every error: it classifies failures, limits attempts and total time, and checks what guarantees the API offers for duplicate requests.

03 Oct 2026 GUIA

Updating an AI SDK Without Breaking Your Integration: Contract Tests

A practical method for checking SDK and API changes, validating observable structures, and separating technical compatibility from the quality of generated responses.

03 Oct 2026 GUIA

Tracing AI Applications: How to Diagnose a Response Without Logging More Sensitive Data Than Necessary

A guide to following a request end to end, finding where a failure occurred, and designing useful telemetry with minimal capture, access controls, and limited retention.

27 Sep 2026 GUIA

Concurrency in AI APIs: How to Control Queues, Quotas, and Latency

A practical method for deciding which requests to accept, defer, or reject before a traffic spike turns into a growing queue, more errors, and unpredictable costs.

24 Sep 2026 GUIA

Streaming AI Responses: How to Show Progress Without Executing Data or Actions Before You Have Valid Output

Streaming improves the perception of speed because it lets you show activity before a complete response exists. However, a received fragment is not yet a business result or a safe instruction. This guide proposes a state machine, a decision matrix, and idempotency controls to separate the provisional interface from persistence, indexing, and actions with external effects.

22 Sep 2026 GUIA

Retries, limits, and duplicates in AI APIs: how to prevent a failure from executing an action twice

A timeout does not prove that an API did not process a request. This guide proposes a design based on identifiable operations, bounded retries, and state reconciliation for integrating models and agents without duplicating emails, purchases, data changes, or tool calls.

22 Sep 2026 GUIA

Contracts for AI Applications: Detecting Model, SDK, or Tool Changes Before Production

A response that appears correct may still violate the format, authorization, provenance, or business rule that supports an AI workflow. This guide explains how to turn those assumptions into verifiable contracts and use them to block regressions before production.

19 Sep 2026

Keep exploring

SourcesHugging Face Transformershuggingface.co