GPT-5.6 Sol reasoning tiers cost the same. Only xhigh is rational.

Flat price, stepped quality

If you are picking a GPT-5.6 Sol (OpenAI) effort tier in July 2026, default to GPT-5.6 Sol (xhigh). Every Sol variant bills $11.25/M tokens, so paying the same rate for medium or high only buys a quieter model and a worse quality index. The ladder is real on benchmarks. The pricing is not.

That is the operational story behind OpenAI’s Sol family. Base GPT-5.6 Sol sits at 58.9 quality. <!-- fact:gpt-5-6-sol|quality=58.9|price=11.25 --> The named effort knobs step down from there: xhigh at 57.7, high at 55.9, medium at 53.6, all still $11.25/M. <!-- fact:gpt-5-6-sol-xhigh|quality=57.7|price=11.25 --> <!-- fact:gpt-5-6-sol-high|quality=55.9|price=11.25 --> <!-- fact:gpt-5-6-sol-medium|quality=53.6|price=11.25 --> Inference latency barely moves. You are not buying a cheaper path when you dial effort down. You are buying a weaker one at full freight.

What the numbers actually say

I lined the three effort SKUs next to the parent Sol checkpoint and looked at quality versus output speed, not marketing labels.

<!-- fact:gpt-5-6-sol|quality=58.9|price=11.25|speed=67 --> <!-- fact:gpt-5-6-sol-xhigh|quality=57.7|price=11.25|speed=64 --> <!-- fact:gpt-5-6-sol-high|quality=55.9|price=11.25|speed=59 --> <!-- fact:gpt-5-6-sol-medium|quality=53.6|price=11.25|speed=60 -->

From xhigh to medium you surrender 4.1 quality points and gain nothing clean on throughput. Medium even edges high on speed (60 vs 59 tok/s) while sitting 2.3 points lower on quality. <!-- fact:gpt-5-6-sol-medium|quality=53.6|speed=60 --> <!-- fact:gpt-5-6-sol-high|quality=55.9|speed=59 --> That is not a latency product. That is a quality haircut with a rounding error on tokens per second.

Translate that into pipelines. A 4-point quality drop on a coding-heavy agent loop means more tool retries, more validator failures, more human review. Retries are where $11.25/M stops being a line item and becomes a multiplier. If structured-output parse failures climb, you pay the Sol rate again on the repair pass. Medium does not make batch jobs cheaper. It makes them longer.

How Sol stacks against the rest of the board

Context matters. Claude Fable 5 (Anthropic) still leads the wider table at 59.9 quality and $20.00/M with 70 tok/s. <!-- fact:claude-fable-5|quality=59.9|price=20.00|speed=70 --> Base Sol is the closest paid chase at 58.9 and roughly half the unit price. <!-- fact:gpt-5-6-sol|quality=58.9|price=11.25 --> xhigh keeps most of that gap closed at 57.7. <!-- fact:gpt-5-6-sol-xhigh|quality=57.7|price=11.25 -->

Drop to medium and the picture flips. Grok 4.5 (xAI) posts 53.8 quality at $3.00/M and 73 tok/s. <!-- fact:grok-4-5|quality=53.8|price=3.00|speed=73 --> Claude Sonnet 5 (Anthropic) is 53.4 at $4.00/M and 86 tok/s. <!-- fact:claude-sonnet-5|quality=53.4|price=4.00|speed=86 --> Medium Sol’s 53.6 at $11.25/M is in that quality band while charging three to four times the rate and running slower. <!-- fact:gpt-5-6-sol-medium|quality=53.6|price=11.25|speed=60 --> For any workload where 53–54 quality is “good enough,” medium Sol is the wrong shop.

High sits in a muddy middle: 55.9 quality, still $11.25/M, 59 tok/s. <!-- fact:gpt-5-6-sol-high|quality=55.9|price=11.25|speed=59 --> Claude Opus 4.8 (Anthropic) is 55.7 at $10.00/M and 64 tok/s. <!-- fact:claude-opus-4-8|quality=55.7|price=10.00|speed=64 --> GPT-5.6 Terra (OpenAI) is 55.0 at $5.63/M with 149 tok/s. <!-- fact:gpt-5-6-terra|quality=55.0|price=5.63|speed=149 --> If you were about to pick Sol high for “balanced” work, Terra is faster and cheaper at similar quality, and Opus 4.8 undercuts on price with a near-match score. High only wins if you are locked into the Sol endpoint for routing reasons.

Why the tiers still exist

I do not think the medium and high SKUs are accidents. Effort knobs are useful when the same checkpoint must serve interactive chat and deep agent traces without a model swap. Product surfaces that expose “thinking level” to end users need more than one stop. Internal A/B harnesses want a controlled quality degradation path.

The counter-argument is latency under load, not the median tok/s on a leaderboard. xhigh at 64 tok/s versus medium at 60 is noise in a single-stream bench. <!-- fact:gpt-5-6-sol-xhigh|speed=64 --> <!-- fact:gpt-5-6-sol-medium|speed=60 --> In a multi-tenant cluster, lower effort can mean shorter critical sections and better tail latency when reasoning traces dominate wall clock. If your p95 is bound by trace length rather than decode rate, medium can still be the right operational choice even when the sticker price matches xhigh.

That case is narrower than the product matrix implies. You need measured trace-length savings large enough to offset more downstream failures. Most code-review, extraction, and tool-calling stacks I see fail the other way: quality shortfalls cost more than the extra tokens in a longer reasoning pass. Flat pricing makes that trade brutal. There is no rebate for dialing down.

Who should use which tier

xhigh is the default for agent loops, multi-step coding, and any pipeline where a wrong structured object triggers a full retry. At equal price to its siblings, 57.7 quality is the only Sol effort number that stays near frontier without jumping to Fable’s $20.00/M. <!-- fact:gpt-5-6-sol-xhigh|quality=57.7 --> <!-- fact:claude-fable-5|price=20.00 -->

high is hard to defend on a clean sheet. 55.9 quality at $11.25/M loses on value to Terra’s 55.0 at $5.63/M and 149 tok/s when throughput matters, and loses on price to Opus 4.8 at similar quality. <!-- fact:gpt-5-6-sol-high|quality=55.9|price=11.25 --> <!-- fact:gpt-5-6-terra|quality=55.0|price=5.63|speed=149 --> Use it only if your router is already Sol-only and you have latency traces proving high beats xhigh on p95.

medium is a trap for cost-sensitive batch. You still pay $11.25/M for 53.6 quality while Grok 4.5 and Sonnet 5 land in the same band for $3–4/M and higher tokens per second. <!-- fact:gpt-5-6-sol-medium|quality=53.6|price=11.25 --> <!-- fact:grok-4-5|quality=53.8|price=3.00 --> <!-- fact:claude-sonnet-5|quality=53.4|price=4.00 --> Medium makes sense as a degraded mode inside a Sol-shaped product, not as a purchase decision from a blank eval.

Base Sol at 58.9 quality and 67 tok/s remains the simple “give me Sol” choice when you do not need an explicit effort pin. <!-- fact:gpt-5-6-sol|quality=58.9|speed=67 --> xhigh is the pinned substitute when the API forces an effort field and you refuse to leave quality on the table.

The pricing design problem

OpenAI shipped a three-rung effort ladder and forgot the part where lower rungs get lower prices. Anthropic and others have played similar games with thinking budgets, but the Sol sheet is unusually blunt: identical $11.25/M across xhigh, high, and medium. <!-- fact:gpt-5-6-sol-xhigh|price=11.25 --> <!-- fact:gpt-5-6-sol-high|price=11.25 --> <!-- fact:gpt-5-6-sol-medium|price=11.25 --> That design teaches one lesson to anyone reading a bill. Never voluntarily select the weaker rung.

Until effort discounts appear, I will keep treating medium and high as compatibility shims, not value tiers. The quality steps are measurable. The speed steps are not. The price step is zero. Choose GPT-5.6 Sol (xhigh) when you are on Sol, or leave the family when 53–56 quality is enough and $5/M-class models clear the bar.

For workload-matched picks across the full field, use the LLM Selector or browse current cards on Explore.