TokenMix Research Lab · 2026-04-21
GPT-5.5 — codenamed "Spud" — finished pretraining on March 24, 2026. No API pricing has been announced. The last five OpenAI releases (GPT-5, 5.1, 5.2, 5.3, 5.4) set a clear pattern, and competitive pressure from Gemini 3.1 Pro at $2/
| Claim | Status | Source |
|---|---|---|
| GPT-5.4 current price: $2.50 input / 5 output per MTok | Confirmed | OpenAI API pricing |
| GPT-5.4 cached input: .25 per MTok (50% discount) | Confirmed | OpenAI docs |
| GPT-5.4 long context (>272K): input doubles to $5 | Confirmed | OpenAI docs |
| Claude Opus 4.7 price: $5 / $25 per MTok | Confirmed | Anthropic launch |
| Gemini 3.1 Pro price: $2 / 2 per MTok | Confirmed | Google DeepMind |
| GPT-5.5 "Spud" pretraining done March 24, 2026 | Confirmed | The Information reports |
| GPT-5.5 pricing above GPT-5.4 | Possible | Depends on capability jump |
| GPT-5.5 pricing same as GPT-5.4 | Likely | Matches GPT-5.1-5.3 pattern |
| GPT-5.5 pricing below GPT-5.4 | Possible | Competitive pressure from Gemini 3.1 |
Bottom line: the pricing band is effectively fixed between $2 (Gemini floor) and $5 (Claude ceiling) for input, and 2-$25 for output. OpenAI cannot price above Claude Opus 4.7 without explicit SOTA benchmarks, and cannot price below Gemini 3.1 Pro without a clear reason to start a price war.
Looking at OpenAI's recent release history reveals a strong bias toward holding price across minor version bumps:
| Release | Date | Input ($/MTok) | Output ($/MTok) | Change vs prior |
|---|---|---|---|---|
| GPT-5.0 | Aug 2025 | $2.50 | $20 | New family |
| GPT-5.1 | Oct 2025 | $2.50 | 8 | Output -10% |
| GPT-5.2 | Dec 2025 | $2.50 | 5 | Output -17% |
| GPT-5.3 | Feb 2026 | $2.50 | 5 | Flat |
| GPT-5.4 | Mar 5, 2026 | $2.50 | 5 | Flat |
| GPT-5.5 (Spud) | TBD | ? | ? | ? |
Sources: OpenAI API pricing history via Wayback Machine, BenchLM GPT-5.4 profile, NxCode GPT-5.4 Guide.
Pattern observation:
What forces a price move? Three triggers historically:
The current frontier pricing landscape:
| Model | Input ($/MTok) | Output ($/MTok) | Blended @ 80/20 I/O | Best feature |
|---|---|---|---|---|
| Gemini 3.1 Pro | $2.00 | 2 | $4.00 | 94.3% GPQA Diamond |
| GPT-5.4 | $2.50 | 5 | $5.00 | 57.7% SWE-bench Pro |
| Claude Opus 4.7 | $5.00 | $25 | $9.00 | 87.6% SWE-bench Verified |
| DeepSeek V3.2 | $0.14 | $0.28 | $0.17 | Dirt-cheap, weaker reasoning |
Blended at 80% input / 20% output, which matches typical coding workload token mix.
What this means for Spud pricing:
Positioning: "Better model, same price." Quiet API update with modest capability gains.
Why this scenario happens:
Probability: 50%. This is the default unless competitive pressure forces a move.
Cost impact: Zero. All existing GPT-5.4 cost models carry over. Budget forecasting unchanged.
What developers should do: Nothing unusual. Swap model ID at release, re-benchmark for quality improvements, update prompts if needed for any new structured output format.
Positioning: "Frontier capability at Gemini 3.1 Pro price." Direct assault on Google's pricing advantage.
Why this scenario happens:
Probability: 30%. Requires OpenAI to break its price stability pattern, but aligns with how they responded to Claude 3.5 Sonnet (GPT-5.1 output cut).
Cost impact: -20% input, -20% output. For a team spending 0K/month on GPT-5.4, monthly bill drops to ~$8K.
What developers should do: Don't prepay credits on GPT-5.4 until Spud price announcement — you'd be locking in a higher rate for a model about to be obsoleted.
Positioning: "Generational leap justifies premium." Spud ships as a meaningful step up, possibly rebranded as GPT-6.
Why this scenario happens:
Probability: 20%. Requires unambiguous SOTA on 3+ benchmarks, which depends on post-training quality (still in progress).
Cost impact: +40% input, +33% output. For a team spending 0K/month on GPT-5.4, monthly bill jumps to ~ 4K at same usage.
What developers should do:
Monthly cost calculations across scenarios, assuming typical coding workload (80% input tokens, 20% output tokens):
Small team — 5M input / 1.25M output per month:
| Scenario | Input cost | Output cost | Total/month | Delta vs GPT-5.4 |
|---|---|---|---|---|
| GPT-5.4 today | 2.50 | 8.75 | $31.25 | — |
| Spud Scenario 1 | 2.50 | 8.75 | $31.25 | $0 |
| Spud Scenario 2 | 0.00 | 5.00 | $25.00 | -$6.25 |
| Spud Scenario 3 | 7.50 | $25.00 | $42.50 | + 1.25 |
Mid-sized product — 500M input / 125M output per month:
| Scenario | Input cost | Output cost | Total/month | Delta vs GPT-5.4 |
|---|---|---|---|---|
| GPT-5.4 today | ,250 | ,875 | $3,125 | — |
| Spud Scenario 1 | ,250 | ,875 | $3,125 | $0 |
| Spud Scenario 2 | ,000 | ,500 | $2,500 | -$625 |
| Spud Scenario 3 | ,750 | $2,500 | $4,250 | + ,125 |
Enterprise scale — 10B input / 2.5B output per month:
| Scenario | Input cost | Output cost | Total/month | Delta vs GPT-5.4 |
|---|---|---|---|---|
| GPT-5.4 today | $25,000 | $37,500 | $62,500 | — |
| Spud Scenario 1 | $25,000 | $37,500 | $62,500 | $0 |
| Spud Scenario 2 | $20,000 | $30,000 | $50,000 | - 2,500 |
| Spud Scenario 3 | $35,000 | $50,000 | $85,000 | +$22,500 |
At enterprise scale, the gap between scenarios is $35K/month — roughly half a senior engineer's salary. Pricing discovery matters.
Three hidden cost traps to watch on Spud release day:
Claude Opus 4.7 shipped with a new tokenizer that produces up to 35% more tokens for the same input text. Per-token price was unchanged, but effective cost jumped 20-35% for identical workloads.
If Spud ships a new tokenizer (plausible given OpenAI's active tokenizer research), your real bill may be higher than the headline suggests. Always benchmark on your actual data within 48 hours of release.
GPT-5.4 doubles input price from $2.50 to $5.00 once context exceeds 272K tokens. If Spud retains this tier structure but extends default context (e.g., to 500K), workloads that previously stayed under the threshold may now cross it.
New model releases typically ship with tighter rate limits for the first 30 days. If your application depends on Tier 4 (current rate: 60K TPM input for GPT-5.4), expect Spud Tier 4 to launch at 20-30K TPM. Plan fallback to GPT-5.4 or Gemini 3.1 Pro for burst traffic.
Five concrete actions before Spud lands:
For step-by-step implementation of the routing and gateway setup, see our GPT-5.5 migration checklist.
No official pricing exists. Based on OpenAI's last 5 releases (GPT-5 through GPT-5.4) and current competitor pricing, three scenarios are plausible: $2.50/ 5 (status quo, 50% probability), $2.00/ 2 (competitive response, 30%), or $3.50/$20 (premium positioning, 20%). The most likely outcome is flat pricing with GPT-5.4.
Possibly. If OpenAI responds to Gemini 3.1 Pro's $2/ 2 pricing, GPT-5.5 could launch at $2.00/ 2 — a 20% cut. This matches the pattern of GPT-5.1's output price cut in response to Claude 3.5 Sonnet. But stability is more common than cuts in recent history.
Only if benchmarks show a clear generational leap. If Spud ships SWE-bench Verified scores above 92% and reclaims multiple SOTA positions, a $3.50/$20 premium is justified. Without that, enterprise customers won't accept a price increase after 8 months of stability.
Almost certainly yes. Claude Opus 4.7 is priced at $5/$25. Even in the premium scenario, GPT-5.5 is unlikely to exceed $3.50/$20. OpenAI has not priced above Anthropic's flagship in any recent release cycle.
If your monthly bill is above $5,000, yes — wait 2-6 weeks for the Spud release and same-day pricing analysis. Below that threshold, commit to GPT-5.4 now and swap when Spud launches through your abstraction layer. Pre-pay no more than 4 weeks of credits.
TokenMix.ai blog publishes same-day pricing comparisons and cost calculators within 24 hours of any major model release, with our GPT-5.5 release tracker updated as new information surfaces.
By TokenMix Research Lab · Updated 2026-04-22