AI Model Cost Comparison 2026: Every Major LLM Priced Per Token

AI model pricing is confusing by design. Input tokens cost different amounts than output tokens. Cached tokens get discounts. Some models charge per image. Batch APIs halve the price but add latency.

Update (September 26, 2026): List prices only tell half the story, because thinking tokens are billed as output and some models think far more than others. At everyday settings, Grok 4.7 and Claude Sonnet 5 cost more per coding task than Claude Opus 5.5 despite lower list prices, and Gemini 3.8 Flash costs more per image than Gemini 3.1 Pro. Our cost calculator and benchmark dashboard now use measured cost per task.

September 2026 made it stranger. The best model on the Artificial Analysis Intelligence Index, Claude Opus 5.5, is no longer the most expensive one. And OpenAI's GPT-6 Sol and Luna match their GPT-5.6 predecessors on quality at half the price.

This guide puts every major model's pricing in one table, then shows you how to pick the right price tier for your workload without overspending.

The Full Pricing Table (September 2026)

Sorted by total cost per 1M tokens (input + output combined):

Quality scores from the Artificial Analysis Intelligence Index v4.3.2. Pricing from Artificial Analysis (first-party APIs), September 2026. AA rebased its index in September 2026, so every score here is lower than in earlier versions of this article.

Two rows stand out. GPT-6 Sol and Claude Sonnet 5 cost exactly the same, and Sol scores 10 points higher. Opus 5.5 costs 40% of Fable 5.1 or GPT-6 Astra and scores 5 points higher than both.

Superseded Models: Don't Pay These Prices

If any of these are still hard-coded in your stack, switching is free money.

Price Per Quality Point

The real question isn't "which is cheapest?" It's "which gives the most intelligence per dollar?"

Cost per quality point is total price per 1M tokens divided by the Intelligence Index score. GLM-5.3 Flash and GPT-6 Luna get you a point for under 2 cents. Muse Spark 1.3 gets you to 48, the same score as GPT-6 Sol, for less than half the money.

How to Estimate Your Monthly AI Cost

Three variables determine your bill:

1. Tasks per month: how many times you call the API
2. Tokens per task: average input + output tokens per call
3. Model choice: the price per million tokens

Formula: (tasks × tokens_per_task × price_per_token) = monthly cost

Example: 200 coding tasks/month × 45,000 tokens each × GPT-6 Luna pricing:

• Input: 200 × 30,000 × $0.10/1M = $0.60
• Output: 200 × 15,000 × $0.50/1M = $1.50
• Total: $2.10/month

The same workload on GPT-6 Sol: $12 + $30 = $42/month. On Claude Opus 5.5: $24 + $60 = $84/month. On Claude Fable 5.1: $60 + $150 = $210/month.

Use our Cost Calculator to run these numbers for your specific workload.

The Tiered Model Approach

The smartest teams don't use one model for everything. They route tasks by complexity:

Tier 1: Hard problems (10-20% of tasks)

• Model: Claude Opus 5.5
• Use for: Architecture decisions, complex debugging, novel problems
• Cost: High per task, but low volume keeps the bill manageable

Tier 2: Standard work (50-60% of tasks)

• Model: GPT-6 Sol, Muse Spark 1.3, or Grok 4.7
• Use for: Feature implementation, code review, writing, daily coding
• Cost: Moderate per task, handles the bulk of work

Tier 3: Simple tasks (20-30% of tasks)

• Model: GPT-6 Luna, GLM-5.3 Flash, or DeepSeek V4.1 Flash
• Use for: Boilerplate, simple edits, email drafts, summaries
• Cost: Fractions of a cent per task

Result: Take the 200-task month above. All on Opus 5.5, it costs $84. Split it 30 tasks on Opus 5.5, 110 on GPT-6 Sol, and 60 on GPT-6 Luna, and it costs about $36. That's 57% less, with Opus still handling the hard work.

Fable 5.1 and GPT-6 Astra don't appear in any tier. Opus 5.5 beats both at 40% of the price. Keep them for the rare task Opus 5.5 fails.

Hidden Costs to Watch

Cache Hit Discounts

Some providers discount cached input tokens:

• Claude Opus 5.5: $0.20/1M for cache reads vs $4.00 for fresh input. That's 95% off.
• Claude Fable 5.1: $0.25/1M for cache reads, down from $1 on Fable 5 (Anthropic pricing).
• OpenAI: Repeated system prompts get cached automatically, and cache reads cost 10% of the input price (OpenAI pricing).

If your workload has repetitive system prompts (agents, batch processing), your effective input cost can land far below the listed price.

Batch API Pricing

OpenAI's Batch API gives 50% off but responses come within 24 hours. Good for:

• Bulk content generation
• Data processing pipelines
• Non-time-sensitive analysis

Not good for: Real-time interactions, coding agents, anything requiring low latency.

Token Efficiency

Some models use more tokens to accomplish the same task. GPT-6 Astra and Claude Fable 5.1 have identical list prices and tie on both Artificial Analysis indexes. But at max effort AA spent about $3.26 per Intelligence Index task on Astra against $7.63 on Fable 5.1, roughly 40%. Same price per token, fewer tokens.

Per-token price is only half the bill. Our benchmark dashboard shows token efficiency per model.

Effort Levels

Reasoning models let you trade tokens for quality. Claude Opus 5.5 scores 58 at max effort, 54 at high, 51 at medium, and 42 at low. Medium is the default, per Anthropic's model overview. Higher effort means more thinking tokens per task, so the score you pay for depends on the setting you pick.

Free AI Models Worth Using

Free models are fine for personal projects and experimentation. For production workloads, GPT-6 Luna ($0.60/1M total) scores 37 and is cheap enough to be effectively free at moderate volume.

When to Upgrade vs Stay Cheap

Stay on budget models when:

• Tasks are repetitive and well-defined
• Quality difference between models isn't noticeable for your use case
• Volume is high (1000+ tasks/day)
• You're building prototypes or MVPs

Upgrade to premium when:

• Error rate on budget models is costing you rework time
• Tasks require reasoning across large contexts
• Output quality directly affects revenue (customer-facing content)
• You're debugging complex, unfamiliar codebases

Which Models Should You Pay For?

The cheapest model that meets your quality bar is the right model. For most developers: GPT-6 Sol ($12/1M total) for everyday work, GPT-6 Luna ($0.60/1M) for volume, and Claude Opus 5.5 ($24/1M) when the task is hard. The tiered example above runs 200 tasks for about $36 a month.

For teams processing high volume: GLM-5.3 Flash or GPT-6 Luna for simple work, Muse Spark 1.3 ($5.50/1M) for complex work. If you need open weights, MiMo-V2.6-Pro scores 46 at $1.30/1M.

Compare all models live in our LLM Leaderboard or AI Model Comparison Chart. For personalized recommendations, try the Model Selector. And for the full "which model is best" breakdown, see our Best LLM in 2026 guide.