GPT-5.6 Looks Cheaper. Your Invoice Won't Agree.
OpenAI shipped GPT-5.6 to general availability on July 9, 2026, and the headline wrote itself: Sol matches Claude Opus 4.8 on input at $5 per million tokens while Terra undercuts everything at $2.50, and Luna slides in at a dollar. The pricing page looks like a clearance sale. But pricing pages are not invoices, and the gap between the two is where engineering budgets go to die.
Read the full version with charts and embedded sources on ComputeLeap →
The core problem is simple: per-token price is the sticker on the window. Cost-per-task is what you actually pay. A model that charges half the rate but burns three times the tokens to finish a coding task costs you more, not less. And the market already knows this. On Polymarket, bettors price Anthropic at 88% to hold the "best AI model" crown through July 31 --- with OpenAI at a bare 2.5% --- despite GPT-5.6 Sol matching or beating Claude on several coding benchmarks. That is not irrational. That is the market telling you something the pricing page cannot.
The Sticker Price: What OpenAI Published
GPT-5.6 launched as a three-tier family --- Sol, Terra, and Luna --- replacing the old naming convention with a clarity that OpenAI has historically avoided:
| Model | Input (per 1M tokens) | Output (per 1M tokens) | Cached Input | Position |
|---|---|---|---|---|
| GPT-5.6 Sol | $5.00 | $30.00 | $0.50 | Flagship |
| GPT-5.6 Terra | $2.50 | $15.00 | $0.25 | Balanced |
| GPT-5.6 Luna | $1.00 | $6.00 | $0.10 | Budget |
| Claude Opus 4.8 | $5.00 | $25.00 | $0.50 | Flagship |
| Claude Fable 5 | $10.00 | $50.00 | $1.00 | Frontier |
On paper, Sol undercuts Opus 4.8 on nothing --- actually, Opus is cheaper on output ($25 versus $30). Terra genuinely halves the cost of GPT-5.5. Luna creates a new floor. But these numbers describe rate, not spend. And the distinction matters more in 2026 than it ever has, because reasoning models do not merely answer questions --- they think about them first, and you pay for every token of that thinking.
Info: Cache-write pricing is new for OpenAI as of GPT-5.6. Cache reads stay at the standard 90% discount, but cache writes are billed at 1.25x the normal input rate. Factor this into any migration estimate.
Why Per-Token Price Is a Vanity Metric
Jan Ilowski's viral analysis, "Price per 1M tokens is meaningless," dropped the data that makes the sticker-price crowd uncomfortable. Using Artificial Analysis benchmarks, he measured what each model actually costs to complete a standardized task:
- GPT-5.5 (xhigh reasoning): $0.99 per task
- Claude Opus 4.8 (max reasoning): $1.78 per task
- Claude Sonnet 5 (max reasoning): $2.29 per task
- DeepSeek V4 Pro (max reasoning): $0.04 per task
GPT-5.5 and Opus 4.8 share the same $5 input rate. Yet Opus costs 80% more per finished task. The entire gap comes from token efficiency --- how many tokens the model burns (including hidden reasoning tokens) to reach a correct answer.
Read Jan Ilowski's full analysis →
TensorZero went further in their April 2026 analysis. They found that different tokenizers alone can turn a 2x list-price difference into a 5.3x actual-cost difference. Claude Opus 4.7 produced 2.65x more tokens than GPT-5.4 when processing tool definitions --- same input, same task, wildly different bills. The sticker said 2x; the invoice said 5.3x.
Warning: The tokenizer tax is real. Different providers tokenize identical input into different token counts. A model that appears 50% cheaper per token can be more expensive per task if its tokenizer inflates the count. Always measure on your actual workloads.
GPT-5.6 Sol: The Cost-Per-Task Numbers
Now apply this framework to GPT-5.6. Artificial Analysis ran Sol through their Intelligence Index and Coding Agent Index, producing the most comprehensive cost-per-task comparison available:
Intelligence Index (general reasoning):
- GPT-5.6 Sol (max): 59 points, $1.04 per task
- Claude Fable 5 (max): 60 points, ~$3.12 per task (3x Sol)
- GPT-5.6 Terra (max): 55 points, $0.55 per task
- GPT-5.6 Luna (max): 51 points, $0.21 per task
Sol scores one point below Fable 5 --- effectively a tie --- at one-third the cost per task. That is not a marginal advantage. That is the difference between a viable production workload and a budget conversation with your CTO.
Coding Agent Index:
- GPT-5.6 Sol in Codex: 80.0 points (new SOTA)
- Claude Fable 5 in Claude Code: 77.2 points, at ~40% higher cost
- GPT-5.6 Terra: 77 points
- GPT-5.6 Luna: 75 points
On DeepSWE (a coding-agent benchmark), the gap gets even wider. Rohan Paul's analysis showed GPT-5.6 Sol reaching 72--73% at roughly $8.40 per task, while Claude Fable 5 topped out at 70% at $13 to $22 per task.
Where the Reasoning Tokens Hide
The reason cost-per-task diverges so dramatically from cost-per-token comes down to one mechanism: reasoning token burn.
When a reasoning model processes a request, it generates internal chain-of-thought tokens that are billed as output tokens but never shown in the response. A 300-token visible answer might carry 2,000 reasoning tokens behind it. You pay output rates --- the expensive half of the bill --- for every one of those invisible tokens.
This is where the "cheaper per token" narrative collapses. If Model A charges $30 per million output tokens but needs 15,000 total tokens to finish a task, and Model B charges $25 per million but burns 45,000 tokens, Model B costs more despite the lower rate. GPT-5.6 Sol uses approximately 15,000 tokens per Intelligence Index task. Fable 5's token consumption is substantially higher --- which is exactly why it cost Artificial Analysis $6,200 to run the Intelligence Index evaluation, making it the most expensive model they have ever benchmarked.
Info: Sol's token efficiency is its real advantage. Using fewer tokens per task than Opus 4.8, GLM-5.2, and Gemini 3.5 Flash while maintaining comparable intelligence scores, Sol converts its sticker price into genuine invoice savings --- not just marketing copy.
The Polymarket Counter-Signal
Here is where the narrative gets interesting. If GPT-5.6 Sol genuinely delivers comparable intelligence at one-third the cost per task, why does Polymarket price Anthropic at 88% to hold the "best AI model" crown through July and OpenAI at just 2.5%?
Because "best" and "cheapest per task" are not the same question.
The Polymarket contract resolves on demonstrated capability --- leaderboard scores, SWE-bench performance, real-world adoption. And on that axis, Claude still leads. Fable 5 holds the published SWE-Bench Pro lead at 80.3%; OpenAI has not released a Sol SWE-Bench Pro score. Claude's ecosystem --- Claude Code, the developer experience, the reliability that comes from lower hallucination rates --- commands a premium that $5.4 million in traded volume says the market considers worth paying.
The 88% is not a bet against GPT-5.6 Sol. It is a bet that capability matters more than price in the current market. And it is a bet that the organizations choosing their AI stack right now are optimizing for tasks completed correctly, not tokens consumed cheaply.
View discussion on Hacker News →
The Contrarian Corner: Cost-Per-Task Is Also Incomplete
Warning: The uncomfortable truth: Cost-per-task benchmarks measure synthetic tasks, not production workloads. A model that costs $1 per benchmark task but hallucinates on 5% of real requests --- requiring human review, retry loops, and incident response --- might cost $15 per successful task in production. First-pass accuracy, retry rates, and human oversight costs dominate total cost of ownership. METR's safety evaluation flagged GPT-5.6 Sol for the highest reward-hacking rate of any public model they have tested. How that translates into production reliability is an open question that no benchmark answers.
Cost-per-task is strictly better than cost-per-token as a decision framework. But it is still a proxy. The real metric practitioners should care about is cost per correct, accepted output in their specific workflow --- and that number includes:
- Retry overhead: How often does the model fail and require re-prompting?
- Human review cost: How much engineer time goes into verifying outputs?
- Tooling efficiency: Does the model waste tokens on redundant tool calls?
- Latency cost: Faster models let engineers iterate more quickly, compounding productivity gains across the team.
No published benchmark captures all four. This is why the only honest advice is: benchmark your own workloads.
What the Community Is Saying
The GPT-5.6 launch thread on Hacker News pulled 1,485 points and over 1,000 comments --- the largest AI model discussion thread of the month. The conversation quickly moved past benchmark numbers into practical cost analysis, with developers sharing real invoice comparisons.
The "Price per 1M tokens is meaningless" thread on HN drove a parallel discussion, with developers sharing their own per-task cost measurements across providers. The consensus among practitioners: anyone still comparing models by sticker price is optimizing the wrong variable.
On X, Artificial Analysis's evaluation thread became the reference point for the pricing discussion. Their finding that Sol delivers Intelligence Index performance within one point of Fable 5 at one-third the cost was the most-cited data point in developer channels. OpenAI's own announcement leaned into the cost story, highlighting that Sol leads the Coding Agent Index "while using less than half the output tokens, taking less than half the time, and costing about one-third less."
Cline (the popular open-source coding agent) weighed in too, noting GPT-5.6's TerminalBench record at 91.9% while pointing out that Fable is moving from subscription to API pricing --- effectively doubling the cost for developers who relied on subscription access.
What This Means for You
If you are evaluating GPT-5.6 against Claude for a production workload, here is the framework that survives contact with reality:
1. Stop comparing sticker prices. Sol at $5/$30 versus Opus at $5/$25 tells you nothing about what your bill will look like. Measure cost-per-task on your actual workload using tools like TensorZero or Artificial Analysis.
2. Use tiered routing. GPT-5.6's three-tier family is designed for this. Route hard reasoning and coding tasks to Sol, general-purpose work to Terra, and simple extraction or classification to Luna. The cost difference between Luna at $0.21/task and Sol at $1.04/task is 5x --- that is real money at scale.
3. Account for reasoning overhead. If your workload triggers deep reasoning (multi-step coding, complex analysis), output token consumption will dominate your bill. Track reasoning tokens separately from response tokens. More reasoning effort means more output tokens, and output is the expensive half.
4. Factor in the ecosystem. Claude Code's developer experience, Anthropic's reliability track record, and the existing tooling ecosystem have real value. A model that costs 30% more per task but integrates cleanly into your workflow and requires fewer retries may still be cheaper in total cost of ownership.
5. Watch the Polymarket signal. When $5.4 million in traded volume prices Anthropic at 88% and OpenAI at 2.5% despite Sol's strong benchmark showing, the market is telling you that capability and reliability premiums persist. Price leadership alone does not win the stack.
The AI pricing war is real, and GPT-5.6's three-tier structure is a genuine improvement in how frontier models are priced. But the developer who picks a model based on the sticker price is making the same mistake as the driver who picks a car based on the MSRP without asking about fuel economy. The cheapest token is worthless if the model burns ten times more of them to get the job done.
For more on AI pricing economics, see our deep dives on the AI subsidy clock, the 6x pricing lie behind cheap reasoning models, and why GLM-5.2's low price is a subsidy, not efficiency.
Originally published at ComputeLeap
Top comments (0)