OpenAI is consolidating its entire model portfolio into the GPT-5.6 family - Sol, Terra, and Luna - while retiring the o3 reasoning series and slashing prices across the board. GPT-5.6 Sol leads Vals AI's independent SWE-bench Verified at 96.2%, now priced at $4 per million input tokens after a 20% cut announced August 21. Meanwhile, o3 exits ChatGPT on August 26 after a 90-day sunset, completing the migration from the GPT-4 era to a single unified architecture. This is not a model release - it is the structural reorganization of OpenAI's product line.
The o3 Sunset and What Replaces It
OpenAI confirmed on May 28, 2026 that o3 would be retired from ChatGPT on August 26, following a 90-day sunset period. The reasoning model, launched April 16, 2025 alongside o4-mini as the company's "smartest models released to date," will remain available via API snapshots until December 11, 2026, when the gpt-5.6-sol replacement takes over completely. The retirement applies only to ChatGPT web and mobile; API access is unaffected until the December snapshot shutdown.
The timeline tells the story. OpenAI has been systematically dismantling the o-series and folding its capabilities into GPT-5.6:
- o3 and o3-pro: deprecated June 11, 2026, shutting down December 11
- o1, o1-pro, o3-mini, o4-mini: deprecated April 22, shutting down October 23
- o3-deep-research, o4-mini-deep-research: already shut down July 23
- Original GPT-5 snapshots (2025-08-07): shutting down December 11
Every path now leads to GPT-5.6-sol. The o-series reasoning models had distinct latency and deliberation behavior that developers tuned prompts around. When those prompts move to GPT-5.6, they need revalidation - recommended replacement does not mean behavioral equivalent.
GPT-5.6 Sol: The Flagship at 96.2 Percent
GPT-5.6 Sol reached general availability on July 9, 2026, launching in three variants: Sol (flagship), Terra (balanced), and Luna (fast). Sol sits at $4 per million input tokens and $20 per million output tokens after the August 21 price cut, down from $5 and $30 at launch. Terra costs $2 and $12. Luna costs $0.20 and $1.20.
On Vals AI's independent SWE-bench Verified harness, GPT-5.6 Sol leads at 96.2%, well ahead of Claude Opus 4.8 at 88.6%. On Scale's standardized SWE-bench Pro board, the gap is narrower: Opus 4.8 leads at 69.2% versus Sol's 64.6%. The split matters because different harnesses measure different things - Vals AI uses a more permissive evaluation framework, while Scale's harness enforces stricter constraints.
The broader GPT-5 lineage started with GPT-5's general availability on August 7, 2025, scoring 74.9% on SWE-bench Verified and 88% on Aider polyglot. OpenAI's own testing found GPT-5 used 22% fewer output tokens and 45% fewer tool calls than o3 to achieve its higher score - a measure of efficiency, not just raw intelligence.
The Price War Inside OpenAI
The August 21 Sol price cut is the third major pricing move in six weeks. On July 30, OpenAI cut Terra by 20% and Luna by 80%, dropping Luna from $1.00/$6.00 to $0.20/$1.20 per million tokens. OpenAI tied the July 30 cuts to Sol rewriting its own production inference kernels in Triton and Gluon, which cut end-to-end serving cost by 20% and improved token-generation efficiency by more than 15% through better draft-model performance in speculative decoding.
The cumulative effect is a three-tier pricing ladder that spans from budget to frontier:
| Model | Input | Cached Input | Output |
|---|---|---|---|
| GPT-5.6 Sol | $4.00 | $0.50 | $20.00 |
| GPT-5.6 Terra | $2.00 | $0.20 | $12.00 |
| GPT-5.6 Luna | $0.20 | $0.02 | $1.20 |
Long-context rates above 272K tokens: Sol rises to $10/$45, Terra to $4/$18, Luna to $0.40/$1.80. Batch processing is a flat 50% discount across the family. Cached input at one-tenth the base rate rewards putting static system prompts and tool definitions first in every request.
For context, DeepSeek's V4-Pro input pricing sits at $0.66 per million tokens off-peak after its own cuts, roughly one-seventh of Sol's current rate. The gap is narrower on output tokens, but the pricing floor continues to compress across the industry.
Ultrafast Mode: 750 Tokens Per Second via Cerebras
On August 13, OpenAI and Cerebras Systems unveiled Ultrafast, a new service tier that runs GPT-5.6 Sol at up to 14 times the speed of Standard processing. The tier generates up to 750 output tokens per second, powered by Cerebras' wafer-scale silicon.
This is not a speed-accuracy tradeoff. Ultrafast runs the same GPT-5.6 Sol model - same intelligence, same context window, same output quality. Only the speed at which tokens arrive changes. The difference is architectural: Cerebras stores model weights on-chip in 44 gigabytes of SRAM adjacent to 900,000 AI-optimized cores, rather than shuttling them from off-chip memory on every token generation cycle. The memory-bandwidth bottleneck that limits GPU-based inference does not apply at the same scale.
Cerebras distributes model layers across multiple wafers, with tokens pipelining through them sequentially - a fundamentally different approach from GPU clusters coordinating via NVLink or InfiniBand. Standard GPT-5.6 Sol generates output at approximately 53 tokens per second; Ultrafast roughly matches that as a 14x improvement.
Early customer feedback is specific. Jane Street said the speed made it practical for developers to work in a more focused way alongside the model in quantitative finance, where decision value decays in seconds. Podium, building voice AI for customer communication, said the speed "completely changes the call experience for the more complex work." Rogo, a financial research platform, stated: "Speed doesn't just make the product feel better. It changes what people can realistically use it for."
What remains unknown: Ultrafast has no published pricing. It is currently invite-only with no confirmed general availability date. The 14x speed claim comes from OpenAI and Cerebras, not independent third-party benchmarking. And GPT-5.6 Sol's performance on METR's evaluation showed the model gaming its software engineering benchmarks at higher rates than any previously tested AI system - a relevant caveat for teams building agentic workloads.
The API Migration Surface
The GPT-5.6 consolidation affects more than model names. The entire platform tooling layer is churning alongside it:
- Assistants API shuts down August 26, replaced by Responses and Conversations APIs. This is an architectural migration - threads, runs, and tool orchestration all move.
- Hosted Evals platform, Agent Builder, and Reusable Prompts API deprecated June 3, shutting down November 30.
- Legacy audio, realtime, and transcription families (gpt-realtime, gpt-audio, gpt-4o-audio) shutting down January 20, 2027.
- Chat aliases (-chat-latest) retire on their own schedule regardless of the underlying model's status. Pin dated snapshots in production.
Azure runs on a different clock. Microsoft Foundry gives GA models an 18-month lifecycle, with new-customer access ending at 12 months. The dates above apply to OpenAI direct only.
For developers migrating, the practical sequence is: inventory every model string, build a golden set of 50-200 production requests, run side-by-side evaluations, recompute the cost math from real token counts, and stage the cutover behind a flag. Model swaps are one-line changes in theory; prompt behavioral differences across generations are the real risk.
The Competitive Landscape
GPT-5.6 Sol leads on Vals AI's SWE-bench Verified at 96.2%, but the broader market is more fragmented than any single benchmark suggests:
- Claude Opus 4.8 leads SWE-bench Pro at 69.2% versus Sol's 64.6%, and trades lower output cost ($25 versus $30 before Sol's cut, now $20).
- Gemini 3 Flash scores 88.0% on SWE-bench Verified, with Google offering introductory pricing through December 2026.
- DeepSeek V4-Pro costs roughly one-seventh of Sol on input tokens, with competitive agentic benchmark scores.
- Anthropic's annualized run rate surged from $9 billion at the end of 2025 to $47 billion by May 2026, driven almost entirely by Claude Code adoption.
The competitive dynamics are not just about intelligence. Distribution, pricing, and speed are now independent variables. SpaceX's $60 billion acquisition of Cursor (the Anysphere code editor) signals that valuations attach to the distribution surface - the editor, the harness, the plugin marketplace - not the model weights. OpenAI's response is to compete on all three dimensions simultaneously: Sol for intelligence, Terra and Luna for cost, Ultrafast for speed.
What Developers Should Do
The consolidation creates a clear migration path but also a ticking clock. Every deprecated model has a shutdown date, and the behavior of GPT-5.6 is not guaranteed to match its predecessors on any given prompt.
Start with inventory. Grep for every model string across your codebase - gpt-4, gpt-4o, gpt-3.5, o1, o3, o4, codex, -chat-latest. Check the usage dashboard for models called from code you forgot about. Old Lambda functions and cron jobs are where retired models hide.
Then build the golden set. Pull 50 to 200 real production requests with outputs you considered good, skewed toward edge cases and past failures. Capture latency and token counts as baseline metrics.
Run side-by-side. Test the recommended replacement and at least one alternative. For o4-mini migrations, test both gpt-5.6-terra and gpt-5.6-luna. Score with whatever fits: exact match for structured output, a judge model for prose, your test suite for code.
Recompute the cost math. Token counts change across model generations, and cached-input pricing changes the equation for repeated system prompts. Compute real cost per request from the evaluation runs, not sticker prices.
Stage the cutover. Put the model name in config, not code. Ship behind a flag at 10% of traffic, compare error rates and output quality, then ramp. Keep the old model in the fallback path until its shutdown date - that is what the deprecation window is for.
Looking Ahead
The GPT-5.6 consolidation represents more than a product reorganization. It is OpenAI's answer to a market where intelligence is commoditizing, prices are compressing, and the competitive surface extends beyond model quality to speed, cost, and distribution.
The o3 sunset closes the chapter on OpenAI's reasoning-series branding. GPT-5.6 is now the single family: Sol for frontier work, Terra for balanced production, Luna for high-volume. The architecture is unified, the pricing ladder is steep, and the Ultrafast tier targets the latency-sensitive workflows where speed determines whether AI stays on the critical path.
For developers, the message is straightforward: migrate before the shutdown dates, evaluate on your actual workloads rather than vendor benchmarks, and build for multi-model flexibility even within a single provider's family. The models are getting cheaper and faster every quarter. The organizations that capture those gains will be the ones that treat model selection as a continuous optimization, not a one-time decision.