From Claude 3's clean three-tier lineup to GPT-5.4's sprawling family — what you actually pay, what the outliers cost, and which model makes sense for your workload.
Choosing an AI model used to be simple. You picked the smartest one you could afford and moved on. In 2026, that logic breaks down fast — because the pricing landscape has fractured into something genuinely confusing, and the "most expensive" option is rarely the most capable.
This post walks through both Anthropic and OpenAI's current lineups with one goal: give you a clear mental model of what you're actually buying at each price point, where the value is, and where the outliers are hiding. No benchmark scores. No marketing language. Just what things cost and why.
We'll cover Claude 3 (the foundation that established the tier structure), Claude 4.5 and 4.6 (the current generation), and then GPT-5 and GPT-5.4 on the OpenAI side — including the models that exist mostly to confuse you.
The most expensive model in the lineup is almost never the right choice. Understanding the tiers is how you stop overpaying.
Claude 3 introduced the Opus / Sonnet / Haiku naming structure that Anthropic still uses today. Understanding it makes every newer model easier to reason about.
The original flagship — the model that made enterprises take Anthropic seriously. Deep reasoning, nuanced instruction following, and a context window that impressed at the time.
The balance model. Strong enough for production use, cheap enough to run at scale. Sonnet became many developers' default because it hit the right point on the quality/cost curve.
The infrastructure tier. Built for high-volume, low-cost tasks: classification, routing, extraction, first-pass summarization. Haiku 3 is still one of the cheapest capable models Anthropic offers.
These are the models that break the pattern. They exist, they're priced as if they're exceptional, and they're mostly a reason to understand why newer models are a better deal.
Anthropic's original fourth-generation flagship. It was priced at a premium when released because it was the best available. That's no longer the case — and the price hasn't moved.
Same price as Opus 4. An incremental update to the generation — improved but not dramatically different. Anthropic kept the same pricing tier rather than discounting it when the 4.5/4.6 generation launched.
Anthropic doesn't discount legacy models — they leave the original pricing in place and release better ones at lower prices. The effect is the same as a price cut, but it pushes users to migrate rather than coast on the old version at a discount.
You'll see the same behavior at OpenAI. Legacy premium models keep their prices; the new generation undercuts them. It's a migration incentive dressed as a pricing strategy.
The 4.5 and 4.6 releases represent Anthropic's strategic reset: match or exceed legacy flagship performance at a third of the cost. The Opus / Sonnet / Haiku structure is preserved, but everything shifts down in price.
The current top of Anthropic's lineup. Deep reasoning, 1M context window, and agentic capabilities that make it well-suited for multi-step workflows. Matches Opus 4.1 on performance at 67% lower cost.
Remarkably flat pricing across multiple Sonnet generations — you're essentially getting a better model for the same $3/$15 rate each time. 40% cheaper than Opus on input, still very capable for most production tasks.
The high-volume workhorse. 4× cheaper than Opus on input, well-suited for routing, classification, extraction, and customer support tasks that don't require deep reasoning.
Anthropic's strategy is simplicity — three clean tiers, no surprises. OpenAI's strategy is breadth — a dozen models across a massive price range.
GPT-5 launched in August 2025 as OpenAI's then-best model. Its pricing is now notably lower than the 5.4 generation — and its 90% cache discount is one of the best in the industry.
Strong frontier model with a 400K context window. Notably cheaper on input than GPT-5.4 ($1.25 vs $2.50), with a best-in-class 90% cache discount that brings effective cached input cost down to $0.125/MTok.
Absurdly cheap for its capability level. At $0.25 input, it undercuts Claude Haiku 4.5 ($1.00) by 4×. Well-suited for straightforward tasks: summarization, classification, basic Q&A, structured extraction.
The cheapest capable model in either lineup. Also supports the 90% cache discount, bringing cached input down to $0.005/MTok. Handles summarization, classification, and light extraction well. Nothing in Anthropic's catalog matches this price.
OpenAI's current generation adds frontier coding and computer-use capabilities directly into the model. Five variants from $0.20 to $180 per million tokens — the widest price spread of any family covered here.
OpenAI's first mainline model to incorporate frontier coding directly. 1M context window — but note: input pricing doubles to $5/MTok above the 272K token mark, making long-context use significantly more expensive.
Near-flagship accuracy at 70% less than GPT-5.4 Standard. Scored 72.1% on OSWorld — nearly matching the full model's 75%. Supports computer use and tool calling. More than 2× faster than GPT-5 Mini.
API-only, no computer use. Built for classification, extraction, and batch jobs where cost-per-call dominates the decision. 4× more expensive than GPT-5 Nano, but meaningfully more capable on current tasks.
12× the cost of GPT-5.4 Standard. Includes a dedicated GPU allocation for ChatGPT Pro subscribers — no shared-compute latency spikes. Designed for legal analysis, medical reasoning, and enterprise workloads where errors are expensive.
Same pricing as GPT-5.4 Standard — it's the same model with interactive reasoning enabled, not a separate pricing tier. The reasoning tokens are billed as output, so complex problems will generate a larger output bill than they appear.
All prices per million tokens (MTok). Outlier rows flagged in amber. Cache discounts vary by model and provider.
| Model | Provider | Input / MTok | Output / MTok | Context | Cache Discount | Best For |
|---|---|---|---|---|---|---|
| Claude 3 Haiku | Anthropic | $0.25 | $1.25 | 200K | 90% off | Ultra-cheap volume tasks |
| Claude 3 Sonnet | Anthropic | $3.00 | $15.00 | 200K | 90% off | General-purpose, legacy |
| Claude Opus 4 | Outlier | $15.00 | $75.00 | 200K | 90% off | Legacy only — migrate away |
| Claude Opus 4.1 | Outlier | $15.00 | $75.00 | 200K | 90% off | Legacy only — migrate away |
| Claude Haiku 4.5 | Anthropic | $1.00 | $5.00 | 200K | 90% off | High-volume production tasks |
| Claude Sonnet 4.6 | Anthropic | $3.00 | $15.00 | 200K | 90% off | Default for most workloads |
| Claude Opus 4.6 | Anthropic | $5.00 | $25.00 | 1M | 90% off | Complex reasoning, long context |
| GPT-5 Nano | OpenAI | $0.05 | $0.40 | 400K | 90% off | Cheapest capable model anywhere |
| GPT-5 Mini | OpenAI | $0.25 | $2.00 | 400K | 90% off | High-volume, moderate complexity |
| GPT-5 | OpenAI | $1.25 | $10.00 | 400K | 90% off | Strong value with cache discount |
| GPT-5.4 Nano | OpenAI | $0.20 | $1.25 | 1M | Varies | Budget tasks, newer architecture |
| GPT-5.4 Mini | OpenAI | $0.75 | $4.50 | 1M | Varies | Dev default — near-flagship quality |
| GPT-5.4 | OpenAI | $2.50$5.00 above 272K | $15.00 | 1M | Varies | Coding, computer use, frontier tasks |
| GPT-5.4 Pro | Outlier | $30.00 | $180.00 | 1M | Varies | Enterprise / dedicated compute |
The single biggest cost reduction available in either ecosystem isn't switching models — it's enabling caching. Anthropic's prompt caching cuts input costs by 90% for cached content across all models. OpenAI's cache discount is 50–90% depending on the model family.
A typical enterprise application with a large system prompt and document context might have 70–80% of its input tokens come from repeated content. At Anthropic's 90% cache rate, that's equivalent to running at a tenth of the listed input price on the majority of your tokens. Stack the Batch API (50% off everything) on top, and you're looking at effective costs that make headline prices almost irrelevant.
The right order of operations: pick the right model tier first, then layer in caching and batch discounts. Most teams skip the second step.
Claude Opus 4/4.1 at $15/$75 and GPT-5.4 Pro at $30/$180 are the two pricing outliers you'll encounter most often. They share a common characteristic: they were priced for a moment in time when they represented the absolute frontier, and the price hasn't come down.
That's not a mistake by Anthropic or OpenAI — it's a deliberate migration incentive. If they discount legacy models, there's less reason to migrate to the new generation. By keeping legacy prices high and releasing better models at lower prices, they push workloads forward without forcing anyone's hand.
The practical implication: if you have active budget on either Opus 4/4.1 or are evaluating Pro tiers, spend 30 minutes running your actual tasks through the current generation models first. The quality difference is rarely worth 3–12× the cost.
Anthropic gives you three clean tiers and a simple pricing philosophy. OpenAI gives you a sprawling family tree with models at nearly every price point from $0.05 to $180 per million tokens. Neither approach is better in the abstract — they're optimized for different decision styles.
What both ecosystems share: the best value is almost never at the top or bottom of the range. The middle tiers — Claude Sonnet 4.6, GPT-5.4 Mini, GPT-5 Mini — are where most workloads belong. The outliers exist for edge cases and legacy compatibility. The budget models are powerful when cost genuinely dominates quality in your specific context.
If you're not sure where your workload belongs, that's exactly the kind of question worth spending real time on before you commit spend. A wrong model choice at scale isn't expensive once — it's expensive every month.
We help business leaders make clear AI infrastructure decisions before committing budget. Not a sales call — a real conversation about what makes sense.