TokenMix Research Lab · 2026-07-28

GPT-6 Release Date 2026: Spud Was GPT-5.5, What's Next?

GPT-6 Release Date 2026: Spud Was GPT-5.5, What's Next?

Last Updated: 2026-07-28 Author: TokenMix Research Lab Data verified: 2026-07-28 - OpenAI model catalog, GPT-5.6 launch materials, OpenAI's Hugging Face security-incident report, Axios, Bloomberg Law, AP, and Reuters reporting

Spud is not GPT-6. OpenAI released Spud as GPT-5.5 on April 23, 2026. The strongest current GPT-6-adjacent evidence is OpenAI's confirmation of an unnamed model more capable than GPT-5.6 Sol. GPT-6 itself still has no official release date, price, API ID, benchmark, or specification.

The rumor changed after July 21. OpenAI disclosed that a security evaluation involved GPT-5.6 Sol and "an even more capable pre-release model." Axios then reported that Sam Altman was heading to Washington to preview OpenAI's most powerful AI yet, while Bloomberg Law described briefings on an upcoming generation of models. Those are real signals. Calling the unnamed system "GPT-6," assigning it an August launch, or publishing benchmark and pricing tables for it remains Speculation.

Table of Contents

Quick Verdict

The direct answer is no: OpenAI has not announced GPT-6, but it has confirmed that a model beyond GPT-5.6 Sol exists in pre-release evaluation.

Claim Status Source
OpenAI has officially announced GPT-6 False as of 2026-07-28 No GPT-6 entry appears in the OpenAI model catalog or company announcements
Spud was the codename for GPT-6 False Axios reported that Spud shipped as GPT-5.5
GPT-5.6 is OpenAI's latest named public model family Confirmed OpenAI GPT-5.6 launch and model catalog
OpenAI evaluated a model more capable than GPT-5.6 Sol Confirmed OpenAI security-incident report
The unnamed pre-release model is GPT-6 Speculation OpenAI did not publish a model name
OpenAI's next system uses persistent groups of parallel agents Likely / reported Axios AM, not an OpenAI specification
A Washington briefing means public release is imminent Speculation GPT-5.6 itself went through a limited government-requested preview before broad release
GPT-6 pricing, context, parameters, and benchmarks are available False as of 2026-07-28 No official product page, pricing row, system card, or API documentation exists
Production teams should migrate now False There is no public GPT-6 API contract to migrate to

What Happened to Spud

Spud stopped being a useful GPT-6 clue on April 23, 2026, when OpenAI released it as GPT-5.5.

Before release, social posts and rumor pages treated the codename as evidence for a larger generational jump. The shipping name settled the question. Axios identified the launch as GPT-5.5, codenamed Spud. Any current article that still presents Spud as an unreleased GPT-6 candidate is carrying an expired rumor forward.

Spud claim What happened Status Practical conclusion
Spud is an internal OpenAI model codename The codename appeared before launch Confirmed historically The codename was real
Spud will become GPT-6 It shipped as GPT-5.5 False Do not use Spud to date GPT-6
Spud proves GPT-6 was trained by April No official source says this Speculation Codename evidence cannot establish training status
Spud's rumored specifications describe GPT-6 No official GPT-6 specification exists Speculation Treat recycled context and parameter claims as invented
A new unnamed pre-release model now exists OpenAI acknowledged one in July Confirmed This is the current signal to watch

This distinction matters for search as well as engineering. The GPT-5.5 Spud release review is the correct page for Spud's shipped product. This article tracks the separate, still unnamed post-GPT-5.6 signal.

GPT-6 Official Status

As of July 28, OpenAI's public model inventory contains GPT-5.6 tiers but no GPT-6 product or API identifier.

Field Public status on 2026-07-28 Confidence Where to verify
Product name GPT-6 Not announced Confirmed absence OpenAI model catalog
Release date Not published Confirmed absence OpenAI company announcements
API model ID Not published Confirmed absence OpenAI model catalog
Input price Not published Confirmed absence No official GPT-6 pricing row
Output price Not published Confirmed absence No official GPT-6 pricing row
Context window Not published Confirmed absence No official GPT-6 documentation
Maximum output Not published Confirmed absence No official GPT-6 documentation
Parameter count Not published Confirmed absence OpenAI has not disclosed one
Benchmarks Not published Confirmed absence No GPT-6 system card or launch page
Pre-release successor to GPT-5.6 Sol Exists, unnamed Confirmed OpenAI incident report

Absence from a catalog does not prove a product will never use the GPT-6 name. It proves only that developers cannot currently build against that name. The operational source of truth is the live model catalog, not screenshots, browser labels, or unverified API error strings.

Evidence Timeline

The timeline has one resolved rumor, one public launch, and one new pre-release signal.

Date Event What it establishes Status
2026-04-23 Spud ships as GPT-5.5 Spud is not GPT-6 Confirmed via Axios
2026-07-07 Reports describe U.S. approval for broad GPT-5.6 rollout after a limited preview Government review can precede release without naming the next model Likely / reported via Reuters syndication
2026-07-09 OpenAI launches GPT-5.6 Sol, Terra, and Luna GPT-5.6 becomes the latest named family Confirmed via OpenAI
2026-07-21 OpenAI discloses evaluation of GPT-5.6 Sol plus a more capable pre-release model A successor-class system exists Confirmed via OpenAI
2026-07-21 Axios describes OpenAI's next model as using large groups of persistent agents Possible product behavior, not a published API contract Likely / reported via Axios AM
2026-07-23 AP reports on advanced model evaluation and a government-vetting framework Safety review is now part of the launch context Likely / reported via AP
2026-07-26 Axios reports Altman traveling to Washington to preview OpenAI's most powerful AI yet Senior-government preview is active Likely / reported via Axios
2026-07-28 Official catalog still shows no GPT-6 No public launch yet Confirmed via OpenAI models

The conclusion is narrow: OpenAI has moved beyond GPT-5.6 in evaluation, but the public evidence does not reveal the commercial name or launch date.

The Four Signals That Matter

Four signals are worth tracking; only one comes directly from OpenAI and confirms the existence of a stronger pre-release model.

Signal Observed fact Claim type What it supports What it does not support
OpenAI incident report Evaluation involved GPT-5.6 Sol and an even more capable pre-release model Confirmed A more capable model exists Name, date, price, architecture
Washington briefing reports Altman planned briefings on next-generation models Likely / reported External review is underway Immediate public launch
Persistent-agent report The next model may run large groups of agents on long tasks Likely / reported Product direction may emphasize agent orchestration Token limits, latency, API shape
GPT-5.6 public release Sol, Terra, and Luna launched July 9 Confirmed Current baseline for price and capability The successor's commercial tier

The first signal is stronger than a leak because OpenAI published it. It is still weaker than a product announcement because the company deliberately left the model unnamed. The other three help describe sequence and context, not a date.

Why Washington Review Is Not a Release Date

A U.S. government preview confirms access by a small audience; it does not establish the public-release day.

OpenAI's own GPT-5.6 preview note provides the closest precedent. A limited set of trusted partners received early access at government request before broader availability. Reuters reporting later described a government-requested delay and a review process around that rollout. The sequence shows that a briefing may be a checkpoint, not a countdown clock.

Government-review signal Safe interpretation Unsafe interpretation Status
Officials receive a model preview The model is mature enough for restricted evaluation Public API launches within days Confirmed first half; second is Speculation
Lawmakers receive capability briefings Policy and risk discussions are active Commercial name is GPT-6 First half Likely / reported; second Speculation
A review window exists Release planning now includes government coordination The model passes automatically Likely / reported
Security researchers test the model OpenAI is collecting capability and safety evidence Benchmarks are final First half Confirmed; second Speculation
GPT-5.6 followed a similar process It is a useful precedent Every later launch follows the same interval Confirmed precedent; extrapolation Speculation

The practical mistake is to turn process evidence into calendar certainty. A review can end in approval, additional mitigations, a narrower release, a renamed tier, or a delay. None can be selected from the public record today.

What the Agent Reports Actually Say

The most interesting reported capability is persistent parallel agent work, but no public specification defines how developers will invoke it.

Axios AM reported that OpenAI's next model can deploy huge batches of agents that work together persistently on complex tasks. That could mean a model-level capability, a product orchestration layer, or a combination of both. The report does not publish an API parameter, agent count, context limit, or per-task price.

Reported concept What is supported Missing evidence Probe status
Multiple agents work in parallel Described by Axios Official product documentation Likely / reported
Work persists on complex tasks Described by Axios Duration, checkpointing, failure semantics Likely / reported
The model itself contains an agent swarm Not established Architecture disclosure Speculation
Developers will receive an agents API parameter Not established API schema Speculation
Parallel agents will be billed as one request Not established Pricing documentation Speculation
The feature will be called GPT-6 Not established Official naming Speculation

For developers, the distinction is expensive. A single request that fans out into 20 internal workers could be priced by tokens, compute time, tool calls, tasks, or a new service unit. Until OpenAI documents billing, "one prompt" cannot be treated as "one model call."

GPT-5.6 Baseline

GPT-5.6 is the only defensible numerical baseline for estimating what a successor would need to beat.

OpenAI launched three GPT-5.6 tiers on July 9: Sol, Terra, and Luna. Their prices and published benchmark results are real GPT-5.6 data. They are not leaked GPT-6 specifications.

Current tier Input / 1M Output / 1M Cache read / 1M Position
GPT-5.6 Sol $5.00 $30.00 $0.50 Highest-capability public GPT-5.6 tier
GPT-5.6 Terra $2.50 $15.00 $0.25 Balanced tier
GPT-5.6 Luna $1.00 $6.00 $0.10 Lowest-cost GPT-5.6 tier
Unnamed pre-release model Not published Not published Not published More capable than Sol in the cited evaluation
GPT-6 Not published Not published Not published No official product

OpenAI states that cache writes cost 1.25 times normal input and cache reads receive a 90% input discount for GPT-5.6. The full GPT-5.6 pricing and benchmark review covers the shipped family. Any successor budget should begin with that baseline rather than a fabricated GPT-6 price.

Pricing Scenarios and Cost Math

The only honest GPT-6 cost model is scenario analysis anchored to GPT-5.6 Sol's confirmed $5/$30 rate.

Monthly workload GPT-5.6 Sol GPT-5.6 Terra GPT-5.6 Luna Formula
10M input + 1M output $80 $40 $16 input_millions x input_rate + output_millions x output_rate
50M input + 5M output $400 $200 $80 Same formula, no cache
100M input + 10M output $800 $400 $160 Same formula, no cache
1B input + 100M output $8,000 $4,000 $1,600 Same formula, no cache

Cost calculation 1: A workload using 10 million input tokens and 1 million output tokens costs 10 x $5 + 1 x $30 = $80 on GPT-5.6 Sol. The same workload is $40 on Terra and $16 on Luna. These are Confirmed calculations using OpenAI's published rates.

Hypothetical successor pricing Input / 1M Output / 1M 10M input + 1M output Status
Same as GPT-5.6 Sol $5.00 $30.00 $80 Speculation scenario
25% premium $6.25 $37.50 $100 Speculation scenario
50% premium $7.50 $45.00 $120 Speculation scenario
Double price $10.00 $60.00 $160 Speculation scenario
Official GPT-6 rate Not published Not published Not calculable Confirmed absence

Cost calculation 2: If an unnamed successor carries a 25% premium over Sol, the same workload would cost $100; at a 50% premium, $120; at double price, $160. These are budget scenarios, not predictions.

Caching changes the baseline materially. For 50 million input tokens with an 80% cache-hit share and 5 million output tokens on Sol, uncached input costs 10 x $5 = $50, cached reads cost 40 x $0.50 = $20, and output costs 5 x $30 = $150. Total: $220, versus $400 without caching.

Sol cache scenario Input cost Output cost Total Saving vs no cache
No cache: 50M input + 5M output $250 $150 $400 $0
50% cache hit $137.50 $150 $287.50 $112.50 / 28.1%
80% cache hit $70 $150 $220 $180 / 45.0%
90% cache hit $47.50 $150 $197.50 $202.50 / 50.6%

Cost calculation 3: An 80% cache-hit share cuts this Sol workload by $180 per month, or 45%. If a future model changes cache pricing, the calculation must be rebuilt from its official rate card.

Benchmark Baseline

There are no GPT-6 benchmark scores; OpenAI's GPT-5.6 table is the current vendor-reported baseline.

Benchmark GPT-5.6 Sol GPT-5.6 Terra GPT-5.6 Luna GPT-5.5 GPT-6
Agents' Last Exam 52.7 50.4 50.3 46.9 Not published
SWE-bench Pro 64.6 63.4 62.7 59.4 Not published
ExploitBench 73.5 52.9 33.2 47.9 Not published
ExploitGym 33.7 23.2 12.4 15.1 Not published
OSWorld 2.0 62.6 50.2 45.6 47.5 Not published
BrowseComp 90.4 87.5 83.3 84.4 Not published

Source: OpenAI GPT-5.6 launch. These are Confirmed vendor claims, not TokenMix independent measurements. Benchmark setup, tool access, reasoning effort, and pass criteria affect comparability.

The useful takeaway is not a projected GPT-6 score. It is the migration threshold: a successor should be tested against Sol on the workloads where Sol already leads, and against cheaper Terra or Luna where price dominates. A larger aggregate score does not guarantee a lower cost per successful task.

Release Window Scenarios

No exact GPT-6 release window can be confirmed; the scenarios below describe evidence conditions rather than probabilities.

Scenario What would need to happen Current support Status
Public release in August 2026 Government review completes, launch page appears, model enters catalog Restricted-preview reporting only Speculation
Public release in September-December 2026 Evaluation continues, mitigations land, product tier is named Compatible with current evidence but not proven Speculation
Release under a non-GPT-6 name OpenAI extends GPT-5.x or creates a task-specific brand Recent Sol/Terra/Luna tiering makes this plausible Speculation
Limited enterprise or government release first Access remains gated after review Consistent with the GPT-5.6 preview precedent Likely, not confirmed for the unnamed model
Release moves into 2027 Safety, product, compute, or policy work takes longer Always possible; no direct evidence Speculation
The current pre-release model never ships in this form Evaluation leads to changes, merger, or cancellation Standard pre-release risk Speculation

The earliest trustworthy release signal will not be a social countdown. It will be one of these: an OpenAI product page, a model-catalog entry, an official pricing row, an API changelog, or a system card. Until one appears, every month-level date is provisional.

API Migration Checklist

There is nothing to migrate to today, but teams can prepare a validation harness without changing production traffic.

Check Current action Launch-day evidence required Risk if skipped
Model ID Keep GPT-5.6 IDs in production Exact ID from official catalog 404 or wrong model
Endpoint compatibility Keep Responses/Chat adapters modular Official endpoint and SDK examples Integration break
Tool schema Snapshot current tool-call tests New JSON schema and error behavior Silent tool failures
Context and output limits Read values at launch Official token limits Truncation and cost spikes
Reasoning controls Feature-detect parameters Official supported values Invalid request errors
Caching Measure current hit rate New write/read multipliers Budget miss
Parallel-agent billing Record task-level usage Official billing unit and usage fields Unbounded spend
Safety behavior Maintain refusal regression set System card and task tests Compliance regressions
Latency Capture p50/p95/p99 baseline Same-region A/B test SLA breach
Quality Use private task suite Blind evaluation vs Sol/Terra/Luna Benchmark-driven downgrade 
def should_enable_unannounced_model(official, eval_result):
    required_docs = {
        "model_catalog_entry",
        "pricing_page",
        "api_schema",
        "usage_fields",
        "system_card",
    }

    if not required_docs.issubset(official):
        return "NO: keep the model disabled outside an isolated evaluation."

    if eval_result["cost_per_success"] >= eval_result["gpt_5_6_sol_cost_per_success"]:
        return "HOLD: the new model did not beat Sol economically."

    if eval_result["p95_latency_ms"] > eval_result["latency_budget_ms"]:
        return "HOLD: quality improved, but the production SLA failed."

    return "CANARY: route 1% of eligible traffic with a hard spend cap."

The GPT-5.6 API access guide is the correct current integration reference. Do not pre-create a gpt-6 string in production and assume it will become valid.

Risks and Falsifiers

The fastest way to keep a prediction honest is to state what would prove it wrong.

Working hypothesis Status Falsifier Action
The unnamed pre-release model is a GPT-5.6 successor Likely OpenAI says it was an evaluation-only branch Remove successor framing
It may become GPT-6 Speculation OpenAI announces a different commercial name Update title and naming table
Public launch may follow government review Speculation Review ends without broad availability Extend the window; do not invent a delay reason
Persistent agents are a launch feature Likely / reported Official docs omit or separate the feature Treat it as a product layer
Pricing may exceed Sol Speculation Official pricing matches or undercuts Sol Replace scenario table with confirmed rates
The model will beat Sol on all benchmarks Speculation Official or independent tests show regressions Publish the regressions
Existing OpenAI-compatible clients will work unchanged Speculation New endpoint or task object is required Build an adapter rather than hiding incompatibility
Risk Likelihood Impact Mitigation
Rumor pages recycle Spud as GPT-6 High Readers inherit a resolved false claim Link the April GPT-5.5 release
A government briefing is mistaken for launch approval High False date certainty Separate access, review, approval, and release
Vendor benchmarks become universal quality claims High Wrong model choice Run private task evals
Agent fan-out creates hidden usage Medium Severe budget overrun Cap tasks, tools, tokens, and wall time
Model name changes before release Medium Broken SEO and code assumptions Track IDs, not rumors
Safety restrictions change after evaluation Medium Behavior drift Re-run refusal and tool tests
Release is gated by region or account Medium Failed rollout plan Verify account access before announcing migration

Final Recommendation

Treat GPT-6 as an unconfirmed commercial name attached to a confirmed technical signal. Spud is settled as GPT-5.5. The July evidence proves that OpenAI is evaluating something beyond GPT-5.6 Sol, but it does not supply a launch date. Keep production on documented models, build a task-level evaluation harness, and wait for an official catalog entry before migrating.

FAQ

Has OpenAI officially announced GPT-6?

No. As of July 28, 2026, OpenAI has not published a GPT-6 launch page, model-catalog entry, API ID, pricing row, or system card. It has confirmed an unnamed pre-release model more capable than GPT-5.6 Sol.

Was Spud the codename for GPT-6?

No. Spud shipped as GPT-5.5 on April 23, 2026. Continuing to label Spud as GPT-6 is a resolved factual error, not an active rumor.

What is the GPT-6 release date?

No official date exists. August, late 2026, and 2027 are all speculation until OpenAI publishes a product page or model-catalog entry. A Washington preview is evidence of evaluation, not a public-release date.

Is the unnamed pre-release model definitely GPT-6?

No. OpenAI confirmed the model's existence but did not disclose its commercial name. It could ship as GPT-6, another GPT-5.x tier, a task-specific product, or not in its current form.

What will GPT-6 cost through the API?

No GPT-6 API price is published. GPT-5.6 Sol currently costs $5 per million input tokens and $30 per million output tokens, which is a usable budget baseline but not a GPT-6 forecast.

Are there real GPT-6 benchmark scores?

No. Tables assigning exact GPT-6 scores to SWE-bench, OSWorld, GPQA, or other evaluations are unverified unless OpenAI or an independent evaluator identifies the tested model. Current numerical baselines belong to GPT-5.6.

Will GPT-6 use multiple AI agents?

Possibly, but this is not an official API specification. Axios reported that OpenAI's next model can use large groups of persistent agents; OpenAI has not documented invocation, billing, limits, or whether the feature belongs to the model or a product layer.

Should developers prepare a GPT-6 migration now?

Prepare tests, not a production migration. Freeze a GPT-5.6 baseline, add model-ID configuration, measure cost per successful task, and require official pricing, schema, usage fields, and safety documentation before enabling traffic.

About TokenMix

TokenMix.ai is an AI API relay that routes OpenAI, Anthropic, Google, DeepSeek, Qwen, and other large language models through one OpenAI-compatible endpoint at https://api.tokenmix.ai/v1. Check the live model catalog, compare current rates on pricing, and use the OpenAI compatibility reference for integration details.

Sources