Coding Agents: The Fork Is at the Model

TL;DR. Claude Code, Cursor, Codex, Cline, and Aider have claimed the coding-agent position. The coding models that ride on top are Opus 4.7, Sonnet 4.6, GPT-5, Gemini 3.1 Pro, and o3. Since most of the products are BYOK (Cursor, Cline, Aider), which model you ride is the heavier decision. Strengths and weaknesses laid out, with the conclusion as scenario-by-scenario pairing recommendations.

Update (2026-08-15) — the product axis held; the model axis turned over twice. This category's call was "the agent choice is narrow and the real split is the model." That call is proven by this piece's own model table aging first — the five products (Claude Code, Cursor, Codex, Cline, Aider) are unchanged, while every model riding on them has changed.

Model in the table below	Current as of 2026-08
Opus 4.7	Opus 5 (2026-07-24) — SWE-bench Verified 96.0%, unchanged $5/$25. Opus 4.8 (05-28) came in between
Sonnet 4.6	Sonnet 5 (2026-06) — the default on Free and Pro
GPT-5, o3	GPT-5.6 (2026-07-09), in three variants: Luna, Terra, Sol
Gemini 3.1 Pro	The Pro tier is still 3.1 Pro (3.5 Pro unreleased); agentic and coding go to 3.5 Flash

Read the per-model strength, price, and benchmark rows below as a 2026-05 snapshot. The current ranking lives in the models overview. The product comparison — BYOK, approval gates, surfaces, extensions — still holds, and that the product axes did not move while the model generation did is the second piece of evidence for this category's call.

The Ground We Call Coding Agents

Definition — a product that runs coding work autonomously, session by session. Whether on CLI, IDE, or cloud, it's the form that gives a model access to a codebase and delegates a unit of work.

The difference from an Agent Framework — a framework is a methodology layered on top of an agent (superpowers, gsd, ecc, gstack), while the agent itself we look at here is the runtime that framework rides on. If a framework is about how to work, an agent is a decision about where to work.

The classification axes are as follows.

Products

Claude Code — Anthropic's first-party CLI

Strengths — Anthropic uses it in-house to write its own code. Directly integrated with 1M-context Opus 4.7, with automatic GitHub Action PR comments now GA. The Skill system, MCP, and subagents are designed model-friendly by the first party.

Weaknesses — model lock-in (Anthropic only). Failure patterns like silent Sonnet downgrade and a Max-plan 200K cap on long-context work accumulate within the same fork. The UI is CLI-core, an entry barrier for GUI-workflow users.

Cursor — the market-share leader among VS Code forks

Strengths — BYOK multi-model (Anthropic, OpenAI, Google, xAI all selectable), the standard for agent-mode IDEs. Public adoption cases at large companies like Stripe and Shopify. Tab completion, multi-file edit, and the composer flow are the benchmark for non-CLI coding workflows.

Weaknesses — Composer's context cutoff — the agent layer truncates below the model's full context window. Pricing-policy changes are a recurring controversy within the quarter. As a VS Code fork, it's dependent on upstream policy changes.

Codex — OpenAI cloud + CLI

Strengths — first-party direct connection to OpenAI models, autonomous background execution in a cloud sandbox + direct GitHub integration. The agentic loop paired with GPT-5 and o3 is well-ordered within the OpenAI ecosystem.

Weaknesses — forced cloud-only is a burden for companies with security policies. No Anthropic or Google models. A recurring position of trailing the Claude 4.x line on coding benchmarks like Aider polyglot and SWE-bench Verified.

Cline — OSS VS Code extension

Strengths — fully BYOK (OpenRouter, Anthropic, OpenAI, local Ollama all possible), with accumulated GitHub stars comparable to first-party products. Tool-use transparency is the ground that earns trust — every tool call, diff, and terminal command is surfaced as an approval step.

Weaknesses — the burden of directly managing models and API keys (setup, refills, rate limits). Not a first party, so it's on the ground of trailing the model companies' SDK changes. The UX is closer to functional.

Aider — OSS CLI, one of the standards of the polyglot benchmark

Strengths — git-aware coding from a single terminal command. The Aider polyglot benchmark is one of the most frequently cited benchmarks for comparing multi-language coding models. BYOK + the lightest setup. Strong on imperative work (fixing a specific function, writing tests).

Weaknesses — no IDE GUI (terminal-centric). Weak agent mode on large codebases — Aider is closer to interactive pair programming than autonomous background execution.

Product comparison table

One-line takeaway — most are BYOK (Cursor, Cline, Aider), and first-party lock-in is the minority (Claude Code, Codex). The deepest fork is which model you ride.

Models

Claude Opus 4.7 — the sober 1M-context flagship

Strengths — opened 1M context at standard pricing. Big refactors and multi-file reasoning run without a context break. Extended thinking is strong on coding-debugging. Top tier on SWE-bench Verified.

Weaknesses — heavy pricing at input $15/1M. The cost-benefit doesn't hold on short, repetitive work. On long-context use, silent downgrade — a failure pattern reported in issues, where the model quietly drops to a smaller context.

Claude Sonnet 4.6 — the coding default

Strengths — a cost-quality balance fit for 80% of coding work. Top tier on both SWE-bench Verified and Aider polyglot. Frequently shows up as the default choice in BYOK agents like Cursor, Cline, and Aider. Input $3/1M.

Weaknesses — a notch below Opus on complex reasoning and 1M context. Opus is more precise on big refactors and multi-file decomposition.

GPT-5 — OpenAI flagship

Strengths — OpenAI Codex / ChatGPT integration, a refined position on tool use and function calling. Stability of tool decisions in the agentic loop. Multi-modal (image and code as joint input).

Weaknesses — a position of trailing Anthropic models on coding-specific benchmarks (Aider polyglot, SWE-bench Verified). Frequent pricing-policy shifts.

Gemini 3.1 Pro — long context + thinking

Strengths — up to 2M context at standard pricing, with thinking mode strong on reasoning. Uniform multi-modal. Google Cloud / Vertex AI integration accelerates large-enterprise adoption.

Weaknesses — a recurring assessment that coding consistency is weak. Integration with Cursor and Cline is also behind that of Claude/OpenAI models.

o3 — deep reasoning

Strengths — top of the benchmarks on math, algorithms, and proofs. Strong on complex debugging and edge-case tracing when given time to think. Top tier on the reasoning-weighted positions of SWE-bench Verified.

Weaknesses — price explosion on big tasks from thinking-token accumulation. Overkill for short, repetitive coding. The agent loop accumulates thinking every turn by design, so latency and cost pile up together.

Model comparison table

One-line takeaway — Sonnet 4.6 holds the default coding position. Opus 4.7 is for big context / complex reasoning, o3 for deep reasoning / debugging, Gemini for the multi-modal / long-context side, GPT-5 for the tool-use position.

Pairing Recommendations

1. Precision refactoring / large codebases — Claude Code + Opus 4.7

The ground that takes in a large codebase in a single breath and runs multi-file changes without a context break. First-party integration + 1M context. The combination Anthropic dogfoods on its own code.

2. Daily coding / fast PR cycles — Cursor + Sonnet 4.6

The most common combination — a BYOK IDE with the default coding model on top. Many large-company adoption cases. The realistic default for price, speed, and quality.

3. Complex debugging / algorithm verification — Aider + o3

A combination of imperative work (debugging a specific function, verifying a proof) with a deep-reasoning model on top. o3 ranks high on Aider polyglot's reasoning-heavy tasks. A strength of the terminal-only workflow.

Pairings that don't work

• Codex + a non-OpenAI model: lock-in makes the combination itself impossible.
• Cline/Aider + GPT-5/Gemini as coding default: possible, but the benchmark data is weaker than the Sonnet 4.6 combination.
• Fully autonomous like Devin, v0, Bolt: high variance in results relative to cost — a signal that recurs in METR autonomy and accumulated HN discussion. The reason they were left out of the main body of this table.

Where the Fast Splits Will Happen

The ground that will split 2026 Q3.

1. First-party lock-in vs. BYOK — with Claude Code bound to Anthropic and Codex to OpenAI, how far does the BYOK market erode? If BYOK takes the market, agent differentiation remains only as UX and integration.
2. The fork of autonomy — Cursor's background agent, Claude Code's GitHub Action, and Devin-style fully autonomous are different autonomy models. Which fork the market converges on is the core signal of the second half of 2026.
3. Repricing of models — Opus 4.7 1M context at standard pricing + Gemini 3.1 Pro 2M context + o3's thinking-token accumulation. The context size × thinking matrix becomes the new axis of price.

These are the grounds where 2026 Q3's market gets carved.

References

Product official

• Anthropic — Claude Code
• Code with Claude 2026 — Managed Agents
• Cursor
• OpenAI — Codex
• Cline GitHub
• Aider — official site + polyglot leaderboard

Benchmarks / evaluation

• SWE-bench Verified
• Aider polyglot leaderboard
• METR autonomy benchmark
• LMArena

User reports / accumulated signals

• Simon Willison
• Ben Kuhn
• Jason Liu
• Paul Gauthier (Aider maintainer — blog/changelog on the official site)
• Hacker News, accumulated Reddit discussion
• OpenRouter usage stats