AI Model Comparison Guide

You open the model picker and see several options. Each has different strengths, context windows, and price points. This guide tells you which model to use for which task, when to switch, and how much it costs.

What You Will Walk Away With

• A clear default model recommendation for each tool
• Decision criteria for when to switch models
• Pricing breakdowns per request type
• A model routing strategy you can use immediately

Quick Selection Guide

By Task

Task	Recommended Model	Why
Complex agentic coding, multi-file refactors	Claude Opus 5	Anthropic’s default Opus; leads Fable 5 on most measured benchmarks at half the price
Tasks larger than a single sitting	Claude Fable 5	Highest-capability tier; its lead grows with task length and complexity
Everyday Claude Code work	Claude Sonnet 5	Recommended cost/performance starting point; account default varies
Cheap parallel / bulk work	Claude Haiku 4.5	Half Sonnet’s introductory token rate; one-third after Sonnet’s promotion
Hardest Codex tasks	GPT-5.6 Sol	Flagship tier for complex coding and professional work
Everyday Codex work	GPT-5.6 Terra	Balanced capability, latency, and cost
High-volume Codex work	GPT-5.6 Luna	Lowest-cost GPT-5.6 tier
Long-running Cursor tasks	Grok 4.5	First-party frontier model for software and broader computer work
Fast iteration (Cursor)	Cursor Composer 2.5	In-house frontier-speed coding model
Large codebase (>200K tokens)	Opus 5, Sonnet 5, GPT-5.6 family, or Gemini 3.1 Pro	1M-class context windows
Multimodal (images, video)	Gemini 3.1 Pro	Strong multimodal support with a 1M context window
Architecture and design	Claude Opus 5	Deep reasoning at high or xhigh effort
Computer use and browser agents	Claude Opus 5	Beats Fable 5 on OSWorld 2.0 at roughly a third of the cost

By Budget

Budget	Primary Model	Alternative
Premium (peak capability)	Claude Fable 5	Claude Opus 5
Standard	Claude Sonnet 5	GPT-5.6 Terra
Speed-focused (Cursor)	Cursor Composer 2.5	Grok 4.5 Fast
Cost-sensitive	GPT-5.6 Luna	Claude Haiku 4.5
Enterprise/Multimodal	GPT-5.6 Sol	Gemini 3.1 Pro

Model Specifications

Full Comparison Table

Model	Provider	Context	Output Limit	Role	Input $/1M	Output $/1M	Speed
Claude Fable 5	Anthropic	1M	128K	Peak Claude	$10	$50	Slower
Claude Opus 5	Anthropic	1M	128K	Default Claude	$5	$25	Standard
Claude Opus 4.8	Anthropic	1M	128K	Previous Opus	$5	$25	Standard
Claude Sonnet 5	Anthropic	1M	128K	Everyday Claude	$2*	$10*	Standard
Claude Haiku 4.5	Anthropic	200K	64K	Low-cost Claude	$1	$5	Fast
GPT-5.6 Sol	OpenAI	1.05M	128K	Flagship	$5	$30	Standard
GPT-5.6 Terra	OpenAI	1.05M	128K	Balanced	$2.50	$15	Faster
GPT-5.6 Luna	OpenAI	1.05M	128K	High-volume	$1	$6	Fastest
Grok 4.5	Cursor / SpaceXAI	500K API**	—	Broad frontier	$2	$6	Standard
Cursor Composer 2.5	Cursor	200K	—	Coding specialist	$0.50	$2.50	Fast
Gemini 3.1 Pro	Google	1M	—	Multimodal	$2***	$12***	Standard

* Sonnet 5 introductory pricing through August 31, 2026; standard pricing is $3 / $15 per MTok afterward.

** SpaceXAI documents a 500K API context for Grok 4.5. The context exposed by a host such as Cursor can be lower, so verify the current picker before a large-context run.

*** Gemini 3.1 Pro prices per tier: $2 / $12 per MTok for prompts up to 200K input tokens, rising to $4 / $18 above that threshold.

Claude Fable 5 (Anthropic)

The highest-capability tier, for work that outlasts a single sitting.

• Released: June 9, 2026
• Context window: 1M tokens with a 128K output limit
• Key strength: Anthropic states Fable 5’s lead over its other models grows with task length and complexity — it is the tier for long autonomous runs, not for routine throughput. Benchmark rankings depend on effort, harness, and task mix, and Opus 5 now leads it on most published measurements.
• Available in: Claude Code v2.1.170+ (/model fable), Cursor (model picker), Claude API (claude-fable-5)

When to use: When the task horizon itself is the differentiator — overnight runs, migrations that span many sessions — select Fable 5 explicitly for the main session. Fable is never the automatic account default, and subagents can inherit the parent model, so pin cheaper subagents to Sonnet or Haiku when cost matters. Before paying 2x, try raising Opus 5’s effort level: low and medium effort on the current generation often beat xhigh on prior models.

Pricing: $10 / $50 per 1M tokens (input/output) — exactly 2x Opus 5. Effort levels run low, medium, high, xhigh, and max. Thinking is always on and cannot be disabled; sending thinking: {"type": "disabled"} returns a 400. Fable 5 requires 30-day data retention and is unavailable to organizations under zero data retention.

Fable 5 is the generally-available, safety-tuned member of the Mythos class — in Anthropic’s words, “a Mythos-class model that we’ve made safe for general use.” Its sibling, Claude Mythos 5, is the same underlying model with safeguards lifted in some areas; initial access is restricted to Project Glasswing cyber defenders and critical-infrastructure providers.

Fable 5 plan access split on July 20

Anthropic settled Fable 5’s billing on July 20, 2026, after several temporary extensions. On Max and Team Premium it is now permanently included in the subscription, capped at 50% of your weekly usage limits — no usage credits, no cutoff date. Pro and Team Standard access it through usage credits at standard API rates instead, and received a one-time $100 credit grant. All models draw from one shared weekly pool, so heavy Opus or Sonnet use reduces your effective Fable headroom.

Claude Opus 5 (Anthropic)

The default Claude tier for complex agentic coding and enterprise work.

• Released: July 24, 2026
• Context window: 1M tokens — both the default and the maximum — with a 128K output limit
• Key strength: Anthropic reports state-of-the-art results on Frontier-Bench v0.1 (more than doubling Opus 4.8 at a lower cost per task) and GDPval-AA v2, roughly 3x the next-best model on ARC-AGI-3, and a win over Fable 5 on OSWorld 2.0 at about a third of the cost. At max effort it lands within 0.5 points of Fable 5’s peak CursorBench 3.2 score for half the cost per task. Its May 2026 knowledge cutoff is the most current of any Claude model
• Available in: Claude Code v2.1.219+, Cursor (model picker), Claude API (claude-opus-5), Bedrock (anthropic.claude-opus-5), Google Cloud, Microsoft Foundry

When to use: Make it the default. Architecture decisions, multi-file refactors, complex debugging, long-running agents, computer use, and knowledge work at scale. Opus 5 is the account default on Max, Team Premium, Enterprise pay-as-you-go, Anthropic API, Claude Platform on AWS, Bedrock, and Google Cloud sessions; Pro, Team Standard, and Enterprise seats default to Sonnet 5.

Pricing: $5 / $25 per 1M tokens (input/output) — identical to Opus 4.8 and half of Fable 5. Fast mode runs at $10 / $50 for roughly 2.5x the output speed, on the Claude API only. Batch processing halves both rates. The minimum cacheable prompt drops to 512 tokens, down from 1024 on Opus 4.8, so shorter prompts now cache with no code change.

Two breaking changes when moving from Opus 4.8

Thinking is on by default. Omitting the thinking parameter meant no thinking on Opus 4.8; on Opus 5 it runs adaptive thinking. Because max_tokens caps thinking plus response text, a request sized tightly around its answer can now truncate mid-response — raise max_tokens or opt out explicitly.

Disabling thinking is capped at high effort. thinking: {"type": "disabled"} combined with xhigh or max effort returns a 400, and the check runs per request. If a route needs thinking off for latency, drop to high or below — or better, leave thinking on and lower the effort level instead.

Claude Opus 4.8 (Anthropic)

The previous Opus, still available — and the fallback target for flagged requests.

• Released: May 28, 2026
• Context window: 1M tokens with a 128K output limit
• Key strength: Around four times less likely than Opus 4.7 to leave flaws in its own code unflagged, with strong long-horizon agentic performance
• Available in: Claude Code, Cursor (model picker), Anthropic API, Bedrock, Google Cloud

When to use: Opus 5 supersedes it at the same price, so new work should start on Opus 5. Opus 4.8 remains worth knowing about for two reasons: it is not deprecated and still serves pinned production traffic, and it is the model that cybersecurity-flagged requests on both Opus 5 and Fable 5 automatically re-run on. It also still supports Priority Tier, which Opus 5 does not.

Pricing: $5 / $25 per 1M tokens (input/output) — unchanged from Opus 4.7. Fast mode runs at 2x the standard rate for 2.5x the speed. Effort defaults to high.

Claude Sonnet 5 (Anthropic)

The cost-effective everyday Claude tier and default on standard subscription seats.

• Released: June 30, 2026
• Context window: Native 1M tokens
• Key strength: A strict improvement over Sonnet 4.6 with a wider cost/performance range; high effort can match an Opus tier on some tasks
• Available in: Claude Code, Cursor, Anthropic API

When to use: Use it for everyday coding, large-codebase analysis, and most cost-sensitive agentic work; increase effort before moving to Opus or Fable. Sonnet 5 is the account default on Pro, Team Standard, and Enterprise subscription seats. Max, Team Premium, Enterprise pay-as-you-go, and Anthropic API sessions default to Opus 5; organization policy can override either mapping.

Pricing: $2 / $10 per 1M tokens through August 31, 2026; $3 / $15 afterward.

Claude Haiku 4.5 (Anthropic)

The cheap, fast tier that powers parallel work.

• Released: October 2025
• Context window: 200K tokens
• Key strength: Fast and inexpensive enough to drive subagents, codemods, and bulk file edits; its $1/$5 rate is half Sonnet 5’s introductory $2/$10 rate and one-third of the later $3/$15 rate
• Available in: Claude Code (subagents and /model), Anthropic API

When to use: Read-only exploration, bulk scans, fan-out subagents, and simple formatting where you don’t need frontier reasoning. The Tier 1 model in a model-routing strategy.

Pricing: $1 / $5 per 1M tokens (input/output).

GPT-5.6 family (OpenAI)

Three durable capability tiers for Codex, ChatGPT Work, and the OpenAI API.

• Released: General availability July 9, 2026 (preview June 26)
• Context window: 1.05M tokens with a 128K output limit for Sol, Terra, and Luna
• Tiers: gpt-5.6-sol for frontier work, gpt-5.6-terra for balanced everyday work, and gpt-5.6-luna for efficient high-volume workloads; the gpt-5.6 alias routes to Sol
• Available in: Codex, ChatGPT Work, and API for all three tiers; standard ChatGPT conversations expose Sol on eligible paid plans

When to use: Sol for the hardest coding, research, design, computer-use, and professional tasks; Terra for the everyday default; Luna for extraction, routing, bulk operations, and latency/cost-sensitive work. In Codex, Free and Go use Terra; Plus, Pro, Business, and Enterprise can choose all three.

Pricing: Sol $5 / $30, Terra $2.50 / $15, Luna $1 / $6 per 1M input/output tokens. Cache reads cost 10% of input and cache writes 1.25× uncached input. Requests above 272K input tokens are billed at 2× input / 1.5× output for the full request.

GPT-5.6 supports none, low, medium, high, xhigh, and max reasoning. Pro is a reasoning mode on the selected GPT-5.6 model, not a separate API model slug. The Responses API also adds persisted reasoning, explicit prompt caching, Programmatic Tool Calling, and multi-agent beta.

Grok 4.5 (Cursor / SpaceXAI)

Cursor’s first-party frontier model for long-running work beyond software engineering.

• Released: July 8, 2026
• Architecture: Mixture-of-experts, jointly trained by Cursor and SpaceXAI
• Available in: Cursor desktop, web, iOS, CLI, and SDK
• Pricing: $2 / $6 per 1M input/output tokens; fast variant $4 / $18

When to use: Difficult, long-running tasks that need creative tool use across software, data science, finance, legal work, or general computer interaction. Composer 2.5 remains the better fit for smaller, speed-focused coding loops.

Grok 4.5 is not a replacement for Composer 2.5. Cursor describes them as different model weight classes, says Composer 2.5 will remain available, and plans more models of that smaller size. Grok is the larger, broader frontier model; Composer is the cheaper coding specialist.

Cursor excluded Grok 4.5 from its launch CursorBench comparison because an older snapshot of the Cursor codebase was accidentally present in training data; do not use that benchmark as independent evidence of model quality.

Cursor acquisition status

On June 16, 2026, SpaceX — not xAI as a standalone buyer — signed a binding agreement to acquire Anysphere, the company behind Cursor, at an implied $60B equity value. The transaction was still subject to closing conditions and regulatory approvals on this page’s verification date; SpaceX expected closing in Q3 2026. xAI had already joined SpaceX in February and the combined AI operation is branded SpaceXAI, but it is inaccurate to say the Cursor acquisition has already closed.

Gemini 3.1 Pro (Google)

A strong multimodal model with a large context window.

• Released: February 2026
• Context window: 1M tokens
• Key strength: Image, audio, and video analysis with Deep Think mode for complex reasoning.
• Available in: Cursor (model picker), direct API

When to use: Tasks requiring more than 200K tokens of context, multimodal analysis (diagrams, screenshots, video walkthroughs), or when you need Deep Think reasoning mode.

Pricing: Tiered by prompt size. $2 / $12 per 1M tokens (input/output) for prompts up to 200K input tokens; above that threshold input doubles to $4 and output rises to $18. Budget for the higher tier whenever you are deliberately using the large context window.

Cursor Composer 2.5 (Cursor)

Cursor’s smaller, lower-cost coding specialist.

• Released: May 18, 2026
• Architecture: Mixture-of-Experts, enhanced with Cursor’s own continued pretraining and reinforcement learning
• Context window: 200K tokens
• Key strength: A substantial step up over Composer 2 — better at sustained work on long-running tasks and more reliable at following complex instructions
• Available in: Cursor and Grok Build

When to use: Fast local iteration in Cursor. Optimized for multi-file edits, code generation, refactoring, and long task chains across hundreds of actions.

Pricing: $0.50 / $2.50 per 1M tokens (standard); $3.00 / $15.00 (fast variant, the default).

Model Routing Strategy

Use this decision tree for day-to-day work:

1. Check the actual account default: Sonnet 5 on standard Anthropic subscription seats, Opus 5 on premium/direct-API accounts and the managed clouds, Sonnet 4.5 on Microsoft Foundry; Terra for everyday Codex work and Sol for the hardest tasks
2. Complex work on the default not landing? Raise Opus 5’s effort level before changing models — high is the default and xhigh suits most agentic coding. This is usually cheaper than a tier upgrade
3. Task horizon spans sessions? Then select Fable 5 explicitly for the main session. Pin subagents to Sonnet/Haiku if cost matters; they may inherit the selected parent model rather than auto-routing down
4. Need frontier long-running work in Cursor? Use Grok 4.5; switch to Composer 2.5 for faster iteration
5. Need budget savings? Use GPT-5.6 Luna or Haiku 4.5 for bulk/parallel work
6. Context exceeds 200K? Use Opus 5, Sonnet 5, any GPT-5.6 tier, or Gemini 3.1 Pro (note Gemini’s price step above 200K)
7. Multimodal analysis? Gemini 3.1 Pro
8. Everything else? Stay with the default

Tip

In Cursor, use Cmd/Ctrl+/ to quickly cycle between models. In Claude Code, use /model to switch and /effort to trade speed for reasoning depth. Match the model to the task — do not pay Opus prices for simple completions. Note that effort is the lever people under-use: low and medium effort on the current generation often beat xhigh on prior models, so sweep effort before you sweep models.

Cost Analysis

Illustrative Cost by Total Tokens

The examples below assume 80% uncached input and 20% output tokens. They are arithmetic illustrations, not measured task averages; caching, reasoning tokens, tool calls, and host pricing change real cost.

Total tokens	Opus 5	Sonnet 5*	GPT-5.6 Sol	Composer 2.5
1K	~$0.009	~$0.0036	~$0.010	~$0.0009
10K	~$0.09	~$0.036	~$0.10	~$0.009
50K	~$0.45	~$0.18	~$0.50	~$0.045
100K	~$0.90	~$0.36	~$1.00	~$0.09

The Opus 5 column also applies to Opus 4.8, which is priced identically. A Claude Fable 5 request costs exactly 2x that column — $10 / $50 per 1M tokens versus $5 / $25 — and Opus 5 fast mode costs the same as Fable 5 at $10 / $50. * The Sonnet 5 column uses its July 2026 introductory rate.

Subscription Context

Cursor Plans

Cursor meters model use against the usage included in your plan and applies the selected model’s current rate. Grok 4.5, for example, is $2/$6 per MTok in standard mode and $4/$18 in fast mode. Plan allowances and available models can change; verify them in Cursor Settings > Usage and on the official pricing page.

Claude Code Plans

Plan	Price	Primary Model	Usage
Pro	$20/month	Sonnet 5 account default; Opus 5 selectable	Lowest paid limits
Max 5x	$100/month	Opus 5 account default; Sonnet/Fable selectable	Higher limits
Max 20x	$200/month	Opus 5 account default; Sonnet/Fable selectable	Highest individual limits

Claude Code’s default is account- and provider-specific: Sonnet 5 on Pro, Team Standard, and Enterprise subscription seats; Opus 5 on Max, Team Premium, Enterprise pay-as-you-go, and Anthropic API sessions. Organization policy can override this mapping. Managed-cloud defaults as of v2.1.219 are Opus 5 on Amazon Bedrock, Google Cloud Agent Platform, and Claude Platform on AWS, while Microsoft Foundry defaults to Sonnet 4.5 and still resolves the opus alias to Opus 4.6. Opus 5 itself requires Claude Code v2.1.219 or later — older versions do not offer it in the picker. Verify both provider and organization settings. Limits change frequently; verify current allowances on Anthropic’s pricing page. Since July 20, 2026, Fable 5 is permanently included on Max and Team Premium at up to 50% of weekly usage limits, while Pro and Team Standard reach it through usage credits.

Codex Plans

Plan	Price	Model	Access
Free / Go	Varies	GPT-5.6 Terra	Entry Codex access
Plus / Pro / Business / Enterprise	Varies	GPT-5.6 Sol, Terra, Luna	Model choice and effort controls

Codex exposes the GPT-5.6 family by plan. Sol is the flagship, Terra is the balanced tier, and Luna is the lowest-cost tier; max reasoning is available to users with GPT-5.6 access.

Performance Benchmarks

Do not read one table as pure model quality

Provider tables mix model settings and agent harnesses. The independent coding index also measures a deployed system — model + Codex, Claude Code, Grok Build, or Cursor CLI — rather than model weights alone. Compare the same benchmark version, effort level, and harness whenever possible.

Provider-reported coding results

Snapshot verified July 11, 2026. These are launch results published by each model provider; -- means the provider did not publish a directly comparable value in the cited release.

Model and setting	SWE-Bench Pro	DeepSWE 1.1	Terminal-Bench 2.1
GPT-5.6 Sol (max)	64.6%	72.7%	88.8%
GPT-5.6 Terra (max)	63.4%	69.6%	87.4%
GPT-5.6 Luna (max)	62.7%	67.2%	84.7%
Grok 4.5 (high)	64.7%	53.0%	83.3%
Claude Sonnet 5 (max)	63.2%	—	80.4%

Cursor’s Grok launch chart separately reported Grok 4.5 versus Composer 2.5 at 64.7% vs 54.0% on SWE-Bench Pro, 83.3% vs 73.0% on Terminal-Bench 2.1, and 62.0% vs 18.0% on DeepSWE 1.0. That supports a large capability gap on these launch runs, but not a deprecation claim: Cursor explicitly continues both models. Cursor also notes that some competitor scores in its chart were self-reported.

Independent snapshot: Artificial Analysis

Artificial Analysis results below were read July 27, 2026. The Intelligence Index is v4.1; the Coding Agent Index is v1.1 and averages DeepSWE, Terminal-Bench v2, and SWE-Atlas-QnA. Artificial Analysis supported Anthropic in evaluating Claude Opus 5 ahead of release, so treat its Opus 5 figures as vendor-adjacent rather than fully arm’s-length.

Model + measured system	Intelligence Index v4.1	Coding Agent Index v1.1	Cost per task
Claude Opus 5 max / Claude Code	61	joint 1st*	$2.03
Claude Fable 5 max + fallback / Claude Code	60	77	$2.75
GPT-5.6 Sol max / Codex	59	80	—
Claude Opus 4.8 max / Claude Code	56	73	$1.80
GPT-5.6 Terra max / Codex	55	77	—
Grok 4.5 high / Grok Build	54	76	—
Claude Sonnet 5 max	53	—	$1.53
GPT-5.6 Luna max / Codex	51	75	—
Composer 2.5 / Cursor CLI	—	52	—

* Artificial Analysis places Opus 5 at xhigh with Claude Code in joint first place on the Coding Agent Index alongside GPT-5.6 Sol with Codex, without publishing a separate index value for it. Opus 5 tops the Intelligence Index at 61, narrowly ahead of Fable 5 at 60, and sets the highest GDPval-AA v2 (1,861 Elo versus Fable 5’s ~1,747) and AA-Briefcase (1,720 versus ~1,574) scores measured so far.

The cost-per-task column is the number that matters most for routing: Opus 5 delivers the top Intelligence Index score at roughly 26% lower cost per task than Fable 5. Opus 4.8 is cheaper still per task, but scores five points lower.

The current Composer 2.5 v1.1 breakdown is approximately 16% DeepSWE, 67% Terminal-Bench v2, and 72% SWE-Atlas-QnA, with the same 52 index for Standard and Fast. Standard averaged about $0.08 and 9.7 minutes per task; Fast about $0.55 and 6.8 minutes. The often-cited Composer score of 62 came from the May v1.0 index, before DeepSWE replaced SWE-Bench-Pro-Hard-AA, and is not directly comparable to these v1.1 values.

Grok 4.5 pairs strong independent coding performance with comparatively low measured cost, but its AA-Omniscience result is a useful warning against overgeneralizing: Artificial Analysis reported 52% accuracy and a 54% hallucination rate. Keep verification in the loop for factual work.

Model Selection Checklist

1. Identify your primary tool: Cursor, Claude Code, or Codex

2. Check the selected/default model: Account/provider-specific in Claude Code (managed-cloud defaults differ), Terra/Sol by Codex workload, or Composer 2.5/Grok 4.5 by Cursor latency and complexity

3. Evaluate task complexity: Simple tasks do not need the most expensive model

4. Check context requirements: Workloads exceeding 200K tokens need Opus 5, Sonnet 5, a GPT-5.6 tier, or another verified 1M-context model

5. Consider budget: Track with /cost (Claude Code), Settings > Usage (Cursor), or Codex dashboard

6. Adjust as needed: Switch models based on task, not habit

Best Practices

1. Default to the best model for tasks that matter — architecture, security review, complex debugging
2. Downgrade for routine work — simple fixes, boilerplate, and formatting do not need an Opus tier
3. Use speed models for iteration — Composer 2.5 in Cursor for rapid trial-and-error cycles
4. Route bulk work to Haiku 4.5 — subagents, codemods, and fan-out scans cost a fraction of Opus
5. Monitor costs weekly — track which models provide the best ROI for your workflow
6. Stay updated — model capabilities and pricing change frequently. Check the Updates page

Note

Model availability and pricing change frequently. Always verify current pricing at the provider’s website before making subscription decisions.

Verification Sources

• OpenAI: GPT-5.6 announcement, model guidance, and Codex changelog
• Cursor / SpaceXAI: official changelog, Cursor’s Grok 4.5 launch, SpaceXAI’s Grok 4.5 release, and Composer 2.5 in Grok Build
• Anthropic: Claude models overview, pricing, Claude Code model configuration, Claude Code changelog, Claude Opus 5, Claude Sonnet 5, and Fable 5 availability
• Independent benchmarks: Artificial Analysis on Claude Opus 5, GPT-5.6, Grok 4.5, live coding-agent leaderboard, and v1.1 methodology
• Google: Gemini API pricing
• Transaction status: SpaceX Form 8-K filed with the SEC