GPT-5.6 Sol Price Cut: $4/$20 Standard, $2/$10 Batch, Locked In Through November 21, 2026

What Changed on August 22, 2026

The August 22 cut was OpenAI's third headline price reduction on a frontier model in 2026 — first GPT-5.5 dropped by 30% in February, then GPT-5.5-pro was repositioned in April, and now GPT-5.6 Sol has come down by 20% on input and 33% on output. The trigger for this particular cut appears to be competitive pressure from two directions: Anthropic Claude Sonnet 5 (now at $3/$15 per million tokens, which became cheaper than Sol at the old $5/$30 price) and the rapid acceleration of the open-weight frontier (Qwen 3.8-Plus, GLM 5.3, DeepSeek V4 Pro all posting near-Sol benchmark scores at one-tenth the API price).

The cut is structured as a promotional discount rather than a permanent price reset. The exact language on the OpenAI pricing page, as of September 3, 2026, is: "Sol's promotional pricing is available at least through November 21, 2026." The "at least" phrasing gives OpenAI the option to extend the cut if competitive conditions warrant, but it also gives them the option to revert to $5/$30 if the market shifts. There is no commitment language about a price floor after November 21.

For budget planning this matters more than the headline numbers. If you commit to a Sol-heavy architecture between now and mid-November, you are betting that either (a) OpenAI extends the promotional window, or (b) the post-November price is at or below the current rate. The safe architectural move is to design for the worst case ($5/$30 revert) and treat any savings below that as upside — that way a price revert does not break your unit economics overnight.

The Four Pricing Tiers After the Cut

OpenAI exposes four pricing tiers for Sol, each with different latency / reliability / cost trade-offs. The August 22 cut applies uniformly across all four tiers in percentage terms, so the relative ordering of the tiers is unchanged. Here is the verified pricing as it appears on the developers.openai.com pricing page on September 3, 2026:

GPT-5.6 Sol — Verified Pricing (2026-09-03)

Tier        Input ($/M)  Cached Input  Cache Write  Output ($/M)
--------    -----------  ------------  -----------  ------------
Standard       $4.00        $0.40         $5.00       $20.00
Batch          $2.00        $0.20         $2.50       $10.00
Flex           $2.00        $0.20         $2.50       $10.00
Priority       $8.00        $0.80        $10.00       $40.00
Long-context   $8.00        $0.80        $10.00       $40.00
  (Standard tier, >272K input tokens)

Three observations on the tier structure:

1. Batch and Flex are priced identically at $2/$10 per million tokens. The difference is operational, not financial: Batch API guarantees completion within 24 hours and accepts async submission, while Flex is synchronous but with no latency SLA — OpenAI may queue or reject Flex requests during peak load.
2. Priority is 2× Standard at $8/$40 per million tokens. Priority is the right pick only when latency is the binding constraint — for example, a real-time voice pipeline where each additional 100ms of TTFT degrades the user experience more than the 2× cost. Most workloads should not be on Priority.
3. Long-context pricing kicks in at 272K input tokens and doubles the Standard rate. This is the same threshold GPT-5.5 used, so existing context-aware architectures carry over unchanged. The 1M-context beta announced alongside the Fable 5.1 launch is not yet available on Sol.

Cache pricing deserves a callout. Cached input dropped from $0.50 to $0.40 per million tokens (20% off, same percentage as input) on Standard, and from $0.25 to $0.20 per million tokens (20% off) on Batch. Cache write pricing dropped from $6.25 to $5.00 on Standard. For prompt-cache-heavy workloads (agentic loops, document Q&A, code review), the cache read price is now competitive with Anthropic Claude Sonnet 5's $0.30 cache read rate — Sol is no longer at a cache-economics disadvantage.

How Sol Compares to Terra and Luna

The GPT-5.6 family has three tiers — Sol (frontier), Terra (mid), Luna (efficient) — and the August 22 cut only changed Sol. Terra stays at $2/$12 and Luna stays at $0.20/$1.20 on the Standard tier. Here is the verified side-by-side after the cut:

GPT-5.6 Family — Standard Tier Pricing (2026-09-03)

Model       Input ($/M)  Cached Input  Cache Write  Output ($/M)  Position
--------    -----------  ------------  -----------  ------------  ---------
Sol            $4.00        $0.40         $5.00       $20.00      Frontier
Terra          $2.00        $0.20         $2.50       $12.00      Mid-tier
Luna           $0.20        $0.02         $0.25        $1.20      Efficient

Sol is 2× more expensive than Terra on input and 1.67× more expensive on output. Sol is 20× more expensive than Luna on input and 16.7× more expensive on output. The capability gap does not match that cost gap on most workloads — Luna scores within 5-8% of Sol on standard benchmarks, and Terra scores within 3-4% of Sol on most reasoning tasks. The case for Sol over Terra/Luna has narrowed with this cut; the previous $5/$30 Sol made the cost-capability ratio hard to justify except for the highest-stakes workloads.

The cut does not change the relative position of Sol within the OpenAI lineup. Sol is still positioned as the recommended default for production workloads that need frontier capability; Terra is the right pick for cost-sensitive workloads where the 3-4% capability gap is acceptable; Luna is the right pick for high-volume workloads where the cost difference overwhelms the capability difference. If you were already on Luna for production, this cut does not change your calculus. If you were on the fence between Terra and Sol for a new project, the cut tips that decision toward Sol on cache-heavy agentic work and toward Terra on simple completions.

The Historical Context: OpenAI's 2026 Price-Cut Pattern

Sol is the third 2026 price cut on an OpenAI frontier model, and the pattern is worth understanding before you build budget assumptions around the November 21 deadline.

• February 2026: GPT-5.5 cut from $2.50/$10 to $1.75/$7 per million tokens (30% off input, 30% off output). The cut was positioned as permanent. It held.
• April 2026: GPT-5.5-pro repositioned at $25/$150 (down from $30/$180) alongside the introduction of GPT-5.5-pro-flex at $15/$90. The cut was permanent and coincided with a new tier introduction.
• August 22, 2026: GPT-5.6 Sol cut from $5/$30 to $4/$20 per million tokens (20% off input, 33% off output). The cut is promotional through November 21.

Two of the three cuts were permanent. The third is promotional with an 11-week window. If you assume a 2/3 probability of permanence and 1/3 probability of revert, the expected price for budget planning between September and December 2026 is approximately $4.30/$22 per million tokens. That is meaningfully better than the $5/$30 baseline and only slightly worse than the $4/$20 promotional tier — a reasonable worst case for capacity planning.

The other thing the pattern tells you: OpenAI is willing to make cuts that reduce the relative price gap to Anthropic. Anthropic Claude Sonnet 5 at $3/$15 is now cheaper than Sol at the new $4/$20 Standard rate, but Sol at the $2/$10 Batch rate is now cheaper than Sonnet 5. If you have a workload that can tolerate async submission or Flex reliability, the Batch tier makes Sol the cheapest frontier API on the market as of September 3, 2026.

What the Cut Means for Different Workloads

The practical impact of the cut depends entirely on your workload's cost structure. Here are the four most common patterns and what changes for each:

1. High-volume text generation (chatbots, content pipelines, RAG). Output tokens dominate cost on most generation workloads, and the 33% output cut is the most valuable part of this announcement. A pipeline that was spending $30,000/month on Sol at the old $5/$30 rate drops to $20,000/month at the new $4/$20 rate — a $10,000/month saving with zero code change. If you can route to Batch or Flex, the saving doubles to $20,000/month against a $30,000 baseline. For most RAG pipelines, Flex is operationally identical to Standard at half the cost.

2. Agentic and coding workloads. Agentic loops are cache-heavy because the system prompt and tool definitions stay constant across turns. The cache read price drop from $0.50 to $0.40 per million tokens matters less than the input price drop for non-cached prompts (20% off), but it matters a lot for the cumulative cache cost on long agent runs. A typical Claude Code-style agent run that touches 50,000 cache reads per session drops from $25 of cache cost to $20 — not transformative but worth tracking if you operate at scale.

3. Long-context (>272K) workloads. The long-context tier did not get the same 33% output cut — it held at the 2× Standard rate ($8/$40). If your workload is long-context-heavy, the relative price advantage of Sol over Terra and Luna narrows considerably. For a 500K-token analysis workload, Sol at $8/$40 long-context is 4× more expensive than Terra at $4/$24 long-context, and the capability gap may not justify the 4× cost. Re-evaluate whether your long-context workload actually needs Sol or whether Terra at the same long-context tier delivers acceptable quality.

4. Latency-critical workloads (real-time voice, live agent UI). The Priority tier also held at 2× Standard ($8/$40). If you were on Priority because your workload needs <500ms TTFT, the price has not changed and your decision matrix is unchanged. If you were on Standard specifically because the absolute price was tolerable, the cut makes the Standard tier's effective cost comparable to what Priority used to cost — but Priority is still 2× faster, so the relative position is unchanged.

API Compatibility and Migration Notes

The model identifier gpt-5.6-sol is unchanged. Existing code that targets Sol continues to work without modification. The price change is transparent — OpenAI bills against the new rates automatically, and the Billing Dashboard surfaces the new rates with no action required from your side.

For workloads that benefit from the cut, the migration is a zero-code-change operation: just confirm your account is on the new pricing tier (most accounts were migrated automatically on August 24) and re-baseline your cost projections. For workloads that want to migrate TO Sol from another provider (Sonnet 5, GPT-5.5, Claude Opus 5), the API surface is OpenAI-compatible, so any OpenAI SDK works with the standard base URL change to https://api.openai.com/v1. The model string changes from claude-sonnet-5 or gpt-5.5 to gpt-5.6-sol and the rest of the request payload is structurally identical.

One operational gotcha: if you have hardcoded the Sol price in your internal cost projections, your projections are now wrong by 20-33% depending on the direction. Audit your internal pricing spreadsheets and update them to reflect the new rates — this is a silent error mode where dashboards show green (costs under budget) but the actual economics are off because the per-token assumption is stale.

Sol Ultrafast Mode: Still Gated

On August 13, 2026 OpenAI and Cerebras Networks published a joint preview of Sol Ultrafast mode — a Cerebras-routed inference path that promises 14× the speed of standard Sol on output tokens. Cerebras's own benchmark claims put Sol Ultrafast at 11× faster than Claude Fable 5 and 5× faster than Claude Opus 4.8 on Fast mode. The HN post on the joint Cerebras / OpenAI launch hit 713 points on August 13 and was the third most-discussed AI API story of August 2026.

As of September 3, 2026, Sol Ultrafast is still gated behind an enterprise sales conversation with no public per-token pricing. HN commenters noted that the absence of pricing on OpenAI's preview page is unusual — "if you have to ask" territory, as one commenter put it. The Cerebras CS-4 wafer-scale silicon that powers the preview is impressive hardware but operates at a cost structure that does not lend itself to commodity pricing.

For workloads that need <200ms TTFT and were considering Sol as a cheaper alternative to Anthropic's premium tiers, Ultrafast is not yet a viable path. Standard Sol on OpenAI's own hardware (H100/B200 clusters) typically delivers 400-800ms TTFT on long prompts, which is acceptable for most production workloads but not for real-time voice or interactive agent UI. Until OpenAI publishes Ultrafast pricing, plan on Standard Sol as the production path and treat Ultrafast as a future capability.

The November 21 Deadline: How to Architect Around It

The "at least through November 21, 2026" language is the operative constraint for any production architecture you build between now and mid-November. Three patterns work well:

Pattern 1: Assume revert, treat the cut as upside. Build your cost projections against $5/$30 Sol Standard and treat any savings below that as bonus. This is the conservative path and the right default for any production system where cost predictability matters more than absolute cost minimization. Your November cost forecast is correct regardless of what OpenAI does on November 22.

Pattern 2: Build an abstraction layer above the price tier. If you operate multiple model providers (OpenAI, Anthropic, OpenRouter, Cloudflare Workers AI), put a price-routing layer in front that can shift traffic to the cheapest equivalent model on a daily or weekly basis. When Sol is at $4/$20, route 80% of your Sol-class traffic to OpenAI Sol and 20% to OpenRouter (also $2/$10 Batch-equivalent). When Sol reverts to $5/$30 in November, the routing layer shifts more traffic to Sonnet 5 at $3/$15 or to OpenRouter's Luna at $0.20/$1.20. The abstraction layer costs ~1 day to build and protects against price volatility.

Pattern 3: Lock in capacity at the promotional rate. OpenAI Enterprise contracts typically allow multi-month price locks at the time of signing. If you have an enterprise relationship, signing a Sol contract that locks the $4/$20 rate through Q1 2027 removes the November 21 uncertainty. This only works at meaningful volume (typically $50K/month+) but is the cleanest answer if you can negotiate it.

Pattern 1 is the default. Pattern 2 is the right answer if you operate at meaningful scale and have engineering capacity. Pattern 3 is the right answer only for enterprise-scale consumers with existing OpenAI relationships.

Sol vs Anthropic Claude Sonnet 5 and Opus 5 (September 2026)

The competitive comparison is the most useful framing for production decisions. Here is the verified side-by-side at Standard tier pricing as of September 3, 2026:

Frontier API Pricing Comparison (2026-09-03, Standard Tier)

Model                Input   Cached Input  Cache Write  Output   Best For
--------             ------  -------------  -----------  ------   -------------------
GPT-5.6 Sol          $4.00      $0.40         $5.00     $20.00   Agentic, coding, tool use
GPT-5.6 Terra        $2.00      $0.20         $2.50     $12.00   Mid-tier, balanced cost
GPT-5.6 Luna         $0.20      $0.02         $0.25      $1.20   High-volume, extraction
Claude Opus 5        $5.00      $0.50         $6.25     $25.00   Premium reasoning
Claude Sonnet 5      $3.00      $0.30         $3.75     $15.00   Balanced production
Claude Haiku 4.5     $1.00      $0.10         $1.25      $5.00   Lightweight production
Gemini 3.7 Pro       $1.25      $0.31         $1.88      $5.00   Multimodal, 1M context
Gemini 3.7 Flash     $0.75      $0.19         $0.94      $3.75   High-volume, multimodal

Sol is now priced between Sonnet 5 and Opus 5 in the frontier tier. Before the cut, Sol at $5/$30 was actually more expensive than Opus 5 at $5/$25 on output — a strange position for OpenAI that this cut corrects. After the cut, Sol at $4/$20 is cheaper than Opus 5 on both dimensions and slightly more expensive than Sonnet 5 across the board.

The capability comparison is more nuanced. On agentic and tool-use benchmarks, Sol and Opus 5 score within 1-3% of each other at standard effort. Sol is meaningfully ahead of Sonnet 5 on hard reasoning chains and on long-context synthesis. Sonnet 5 is meaningfully ahead of Sol on latency and on cost predictability (Anthropic has not done promotional pricing on Sonnet 5 in 2026). For most production workloads, the choice between Sol and Sonnet 5 is a wash on capability and comes down to which API surface your team prefers and whether you want to wait for the November 21 deadline to resolve.

Final Verdict

The August 22 GPT-5.6 Sol price cut is a real improvement for production workloads that can ride the Standard tier through November 21, 2026. The 20% input cut and 33% output cut are the largest percentage moves OpenAI has made on a frontier model in 2026, and the cut positions Sol as the second-cheapest frontier API behind only OpenRouter's promotional Luna tier. For workloads that can route through OpenRouter or to the Batch / Flex tier, the effective price drops to $2/$10 per million tokens — the same rate as the open-weight frontier (Qwen 3.8-Plus, GLM 5.3) sits at on Cloudflare Workers AI.

The cut is promotional, not permanent. Architect for the possibility of a November 22 revert to $5/$30 Standard. Build abstractions if you can. Sign enterprise contracts if you can. And audit any internal cost projections that hardcoded the old rate — they are silently wrong as of August 24, 2026.

For teams on Luna or Terra today, the cut does not change the calculus — Sol is still meaningfully more expensive and the capability gap does not justify the cost difference for most workloads. For teams on Opus 5 or Sonnet 5, the cut is a real opportunity to migrate agentic and coding workloads to Sol at a competitive price point, with the November 21 deadline as the architectural constraint. For teams on the open-weight frontier (Qwen, GLM, DeepSeek), the cut does not threaten the cost advantage — those models are still 5-20× cheaper than Sol even after the cut.