Models

AI Models for Coding in 2026: Every Claude & GPT Model Compared

Four current Claude models, four current GPT models, one table. Prices, context windows, and benchmarks verified July 18, 2026, and a straight answer on which to use for what.

By Kylian Migot · Updated July 2026 · 8 min read

Quick answer

The current lineup: Anthropic ships Claude Fable 5 (frontier), Opus 4.8 (flagship), Sonnet 5 (workhorse), and Haiku 4.5 (fast); OpenAI ships GPT-5.6 in three tiers, Sol, Terra, and Luna, alongside the previous flagship GPT-5.5. The short answer: default to Opus 4.8 or Terra for everyday coding, and reach for Fable 5 or Sol when the work is genuinely hard.
Data verified
July 18, 2026
Claude lineup
Fable 5 · Opus 4.8 · Sonnet 5 · Haiku 4.5
OpenAI lineup
GPT-5.6 Sol · Terra · Luna · GPT-5.5
Everyday defaults
Opus 4.8 or Terra; Fable 5 or Sol for the hardest work
01

Every current model, one table

Every generally available Claude and GPT coding model as of July 18, 2026, with pricing from vendor docs and headline benchmark scores from the sources cited in each row. Model names link to our deep-dive pages where we have them.

ModelVendorPriceContextTierReleasedHeadline benchmark
Claude Fable 5Anthropic$10 in / $50 out per MTok1M tokens (128k output)frontierJune 9, 2026SWE-bench Verified: 95.0%
Claude Opus 4.8Anthropic$5 in / $25 out per MTok (Fast mode $10/$50)1M tokensflagshipMay 28, 2026SWE-bench Verified: 88.6%
Claude Sonnet 5Anthropic$3 in / $15 out per MTok (Intro pricing $2/$10 through Aug 31, 2026)1M tokens (default)workhorseJune 30, 2026
Claude Haiku 4.5Anthropic$1 in / $5 out per MTok200k tokensfastOctober 2025
GPT-5.6 SolOpenAI$5 in / $30 out per MTokNot publishedfrontierJuly 9, 2026Terminal-Bench 2.1: 88.8%
GPT-5.6 TerraOpenAI$2.5 in / $15 out per MTokNot publishedworkhorseJuly 9, 2026Terminal-Bench 2.1: 87.4%
GPT-5.6 LunaOpenAI$1 in / $6 out per MTokNot publishedfastJuly 9, 2026Terminal-Bench 2.1: 84.7%
GPT-5.5OpenAI$5 in / $30 out per MTokNot publishedflagshipApril 23, 2026SWE-bench Verified: 88.7%
02

The two lineups explained

Anthropic and OpenAI structure their families differently, and misunderstanding the naming is how people end up on the wrong model. Anthropic uses named tiers: Fable is the frontier tier (introduced June 9, 2026 as the first Mythos-class model above Opus), Opus is the flagship, Sonnet the mid-tier workhorse, Haiku the fast/cheap tier. Higher tier means more capability and more dollars per token.

OpenAI's GPT-5.6 family works on two axes: the number is the generation, the name is the capability tier. Sol, Terra, and Luna are all GPT-5.6-generation models released the same day (July 9, 2026); Sol is the flagship and the only tier with max reasoning effort and ultra mode, Terra is the workhorse, Luna the budget tier. GPT-5.5 is the previous generation's flagship, still available, and per OpenAI's launch materials Terra now roughly matches it at about half the price.

Claude: tiers by name

Fable 5 (frontier, $10 in / $50 out per MTok) → Opus 4.8 (flagship, $5 in / $25 out per MTok) → Sonnet 5 (workhorse, $3 in / $15 out per MTok) → Haiku 4.5 (fast, $1 in / $5 out per MTok). One name per tier, priced accordingly.

GPT-5.6: generation + tier

Sol (flagship, $5 in / $30 out per MTok) → Terra (workhorse, $2.5 in / $15 out per MTok) → Luna (fast, $1 in / $6 out per MTok), all generation 5.6. GPT-5.5 ($5 in / $30 out per MTok) is the prior generation, superseded as flagship by Sol.

Ship your first agent today

Download AIDEN free and point it at your existing Claude Code or Codex setup. No credit card, running in minutes.

Download AIDEN free

Free to start · macOS 12+ · No credit card required

03

How to choose: match the model to the task shape

Stop asking which model is best and start asking what shape the task is. The same four shapes come up in every codebase, and each has a clear pick in both lineups:

Hard architecture and refactors

Cross-cutting changes, gnarly migrations, long autonomous runs where a failed attempt costs hours. Pick Claude Fable 5 or GPT-5.6 Sol, the frontier premium is cheap next to your time.

Everyday feature work

Implementing stories on a real codebase, review passes on big diffs, normal agentic sessions. Opus 4.8 (Anthropic's own recommended starting point for complex agentic coding) or Terra. This is where most tokens should go.

Volume and cost pressure

Routine features, tests, refactors, or big parallel agent fleets where the bill dominates. Sonnet 5 (Intro pricing $2/$10 through Aug 31, 2026) or Luna. Sonnet 5 keeps a 1M-token context at mid-tier pricing.

Subagents and mechanical edits

Fan-out subagents, classification, triage, lint-fix-grade changes. Haiku 4.5 or Luna, paying flagship prices here is burning money.

For the within-vendor decision in detail, see the best Claude model for coding and the best Codex model. For the head-to-head at the top, Fable 5 vs Sol.

04

Benchmarks, honestly

Read vendor benchmark charts with one rule in mind: each vendor publishes the benchmark it wins. Anthropic leads SWE-bench Verified, where Fable 5 scores 95.0% per the swebench.com leaderboard, verified July 18, 2026. OpenAI leads Terminal-Bench 2.1, where Sol scores 88.8% per OpenAI's launch materials (91.9% in ultra mode). Both claims are true. Neither settles anything on its own.

The most useful cross-vendor signal right now is SWE-bench Pro, because the Sol number comes from third-party runs rather than OpenAI's marketing: Fable 5 scores 80.3% (Anthropic / Scale leaderboard) against 64.6% for Sol (third-party runs (not OpenAI-published)). That is a real gap on repository-level software engineering, and it is the main reason we treat Fable 5 as the strongest coding model available today.

05

Use them together, not either/or

The teams getting the most out of these models do not pick a side. They run Claude models through Claude Code and GPT-5.6 models through Codex CLI, and route each task to whichever model fits its shape and budget, Fable 5 on the migration, Terra on the CRUD endpoints, Haiku on the subagents.

That routing is exactly what AIDEN is built for: a macOS desktop app that orchestrates your existing Claude Code and Codex CLIs on one kanban board. Each story card gets an AI-drafted spec you approve, its own git worktree, and whichever agent, Claude or Codex, you assign to it, running locally on the keys you already pay for (BYOK, free for one project). See how the workflow ranks against the field in the best agentic IDEs of 2026.

FAQ

What is the best AI model for coding right now?
For most day-to-day engineering, Claude Opus 4.8 ($5 in / $25 out per MTok) or GPT-5.6 Terra ($2.5 in / $15 out per MTok) are the right defaults: strong capability at workhorse prices. Reserve Claude Fable 5 or GPT-5.6 Sol for genuinely hard work, cross-cutting refactors, architecture, long autonomous runs. Data on this page was verified July 18, 2026.
Is Claude Fable 5 worth twice Opus 4.8's price?
Sometimes. Fable 5 costs $10 in / $50 out per MTok versus $5 in / $25 out per MTok for Opus 4.8, and it leads every published SWE-bench variant (95.0% on SWE-bench Verified, per the swebench.com leaderboard). For hard multi-step work where one failed run costs you an afternoon, the premium pays for itself. For routine features and fixes, Opus 4.8 delivers most of the capability at half the price.
What is the difference between GPT-5.6 Sol, Terra, and Luna?
They are three capability tiers of the same GPT-5.6 generation, released July 9, 2026. Sol ($5 in / $30 out per MTok) is the flagship and the only tier with max reasoning effort and ultra mode. Terra ($2.5 in / $15 out per MTok) is the everyday workhorse, roughly GPT-5.5 performance at about half the price. Luna ($1 in / $6 out per MTok) is the budget/fast tier for quick mechanical tasks.
Do coding benchmarks actually matter?
Directionally, yes; precisely, no. Each vendor publishes the benchmark it wins: Anthropic leads SWE-bench Verified, OpenAI leads Terminal-Bench 2.1. Third-party SWE-bench Pro runs put Claude Fable 5 well ahead of GPT-5.6 Sol, but no benchmark includes your codebase, your test suite, or your conventions. Use benchmarks to shortlist, then run both models on a real story from your backlog.
Can I use Claude and GPT models together?
Yes, and mixing per task is often the best setup: Claude models via Claude Code, GPT-5.6 models via Codex CLI. AIDEN, our macOS desktop app, orchestrates both CLIs on one kanban board, so you can assign a Claude agent to one story and a Codex agent to another, each in its own git worktree, using the keys and subscriptions you already pay for.

Keep reading

Run any of these models, in parallel.

AIDEN orchestrates your Claude Code and Codex CLIs on one kanban board, so each story gets the right model. Free for one project.

macOS 12+ · Bring your own Claude Code or Codex · Your code stays local