GPT-5.5 and the Thinking Effort Controls: What Actually Changed in April 2026

If you opened ChatGPT in the last week of April 2026 and reached for the model picker looking for an “o4” or “o5,” they were not there – because they never existed. What did land was GPT-5.5 on April 23, a redesigned model picker on April 28 with thinking-effort controls, and a quiet retirement of every named o-series reasoning model. Here is the actual map of what shipped, what got renamed, and how to use the new controls in both ChatGPT and the API.

The headline change is not “a new model.” It is a structural one: OpenAI has finished folding the o-series into the main GPT-5 family and replaced the named-model selector with an effort dial. For practitioners, the implication is concrete – the way you describe “which model to use for hard problems” is now a parameter, not a vendor SKU. The rest of this post walks through what actually shipped, the parts of the docs that are easy to misread, and the workflow rules worth building into your defaults.

The April 2026 shake-up: three changes that arrived in two weeks

Three related changes hit in the last week of April, and they only make sense in combination.

First, GPT-5.5 launched on April 23, 2026 in ChatGPT (Plus, Pro, Business, Enterprise) and Codex, with API access following on April 24 (OpenAI announcement, CNBC). OpenAI president Greg Brockman called it “a new class of intelligence for real work.” It is the first fully retrained base model since GPT-4.5 – the 5.1 through 5.4 series were refinements on the same weights (The Next Web).

Second, the model picker moved into the message composer on April 28, 2026, exposing the new thinking-effort toggle for any GPT-5.x Thinking model. The old top-bar picker that surfaced an obscure “ChatGPT” menu is gone (TechRadar, OpenAI Help Center).

Third, the o-series is now retired from ChatGPT. OpenAI o4-mini went on February 13, 2026 alongside GPT-5 (Instant and Thinking); OpenAI o3 was retired on May 28, 2026 with a 90-day sunset that ends August 26, 2026 (Model Release Notes). If you were routing calls through o3 by name in production, you have until late August to migrate to gpt-5.5 or gpt-5.4.

These three are not separate news items. They are the same change expressed in three surfaces: the model, the interface, and the naming convention. The rest of the post breaks down each surface.

The new ChatGPT lineup, end to end (Instant, Thinking, Pro, and the hidden dials)

ChatGPT today ships three capability tiers and four hidden effort options – and the public-facing label does not always match the model under the hood.

The capability tiers are:

• Instant – the fast, default tier. Despite the name, this is currently GPT-5.3 Instant on most user sessions, not GPT-5.5. It is the speed-optimized path for high-frequency, lower-complexity work (TechRadar).

• Thinking – the reasoning-capable tier. GPT-5.5 Thinking for Plus, Pro, Business, Enterprise; GPT-5.4 Thinking remains available as the professional workhorse tier below it (OpenAI Help Center).

• Pro – the highest-capability variant, available only to Pro/Business/Enterprise. Uses parallel test-time compute. Apps, Memory, Canvas, and image generation are not available with Pro.

The hidden controls are the part the docs do not surface prominently. When you select GPT-5.5 Thinking or GPT-5.4 Thinking in the model picker, a thinking-time toggle appears in the message composer (OpenAI Help Center):

• Standard (new default for Plus/Business) – a lightweight reasoning pass, faster than the old default.

• Extended – the previous default for Plus; full reasoning process with self-correction.

• Light (Pro only) – the snappiest reasoning option.

• Heavy (Pro only) – the deepest reasoning tier; responses can take 30-90 seconds on genuinely hard problems.

Auto-switching from Instant to Thinking is on by default. When you select Instant, ChatGPT can route a complex request to Thinking automatically without burning your manual Thinking budget. You can disable auto-switching in Configure in the model picker if you want predictable billing.

One subtle point: the public-facing “Instant” tier is GPT-5.3 Instant, not GPT-5.5. If you specifically need GPT-5.5’s capabilities at speed, you have to select Thinking and then drag the effort down to Light or Standard. This is the single most common point of confusion in the new picker.

What the model picker actually does under the hood

The picker is a UX wrapper over OpenAI’s unified model system. Since the launch of GPT-5 in August 2025, every GPT-5.x model has been a configuration on the same underlying system: a fast mainline model, a deeper thinking model, and a real-time router that decides which to use (OpenAI GPT-5 system card, Introducing GPT-5). What changed in April 2026 is that the router became user-facing.

In the GPT-5 system card, OpenAI spelled out the mapping that the picker now exposes:

Old (retired)	Current equivalent	What changed
GPT-4o	GPT-5.4 Thinking	Significantly stronger, reasoning-native
GPT-4o mini	GPT-5.3 Instant / GPT-5.4 mini	Faster, cheaper
o1	GPT-5.5 Thinking (Thinking effort)	Integrated, no separate model selector
o1-pro	GPT-5.5 Thinking (Extended)	Now accessible via effort controls
o3	GPT-5.5 Pro	Folded into unified tier

That last row is the punchline. There is no “o5” because there is no longer a separate o-series. The reasoning capability you used to access by selecting “o3” is now an effort setting on the main model family.

For comparison context, see our Claude Opus 4.7 vs GPT-5.4 vs Gemini 3.1 Pro frontier benchmark – that piece was written just before GPT-5.5 launched and the relative ordering has shifted. Terminal-Bench 2.0 in particular moved materially: GPT-5.5 now leads at 82.7%, with Opus 4.7 at 69.4% and Gemini 3.1 Pro at 68.5% (Vellum benchmark table).

The API side: reasoning_effort and how to set it for GPT-5.5

If you are calling the API rather than ChatGPT, the picker change translates into a single parameter. Use the Responses API with `reasoning: { effort: “low” | “medium” | “high” | “xhigh” }` ([Reasoning models API guide](https://developers.openai.com/api/docs/guides/reasoning)).

from openai import OpenAI
client = OpenAI()
response = client.responses.create(
model="gpt-5.5",
input=[
{"role": "user", "content": "Plan a refactor of this Python service into three modules."}
],
reasoning={"effort": "high"},
text={"verbosity": "medium"},
)
print(response.output_text)

The four effort levels map to four use patterns:

• low – efficient reasoning with a modest latency increase. Recommended for voice, fast information retrieval, and classification.

• medium – the default for GPT-5.5. Best balance of quality, reliability, and performance; the documented starting point.

• high – hard reasoning, complex debugging, deep planning, agentic tasks. Recommended for most production agent workflows.

• xhigh – deep research, asynchronous workflows, long-running agentic tasks. Only when your evals show a clear benefit; this materially increases latency and token cost.

A few API details worth keeping in your head:

1. Defaults are model-dependent, not universal. GPT-5.5 defaults to medium ; GPT-5.6 defaults to medium in both standard and pro reasoning modes. Always check the model page before assuming (Reasoning API docs).

1. Reasoning tokens are billed as output tokens. You will see them on the API invoice. They are not free (Stob pricing analysis).

1. reasoning.mode (standard vs pro) is independent of reasoning.effort . Mode selects the execution path (standard inference vs parallel test-time compute); effort controls how much reasoning that mode applies. If you want GPT-5.5 Pro-level reasoning in the API, you select the gpt-5.5-pro model SKU, not a mode parameter.

1. Prompts over 272K input tokens get priced at 2x input and 1.5x output for the full session, including GPT-5.5. Long-context work is not a free upgrade (GPT-5.5 API model docs).

1. For latency-sensitive work, set a low effort and consider asking the model for a short preamble first. This gets the visible response started faster.

What the benchmarks actually say (and where the limits are)

The official GPT-5.5 launch numbers are real and meaningful, but they are not the whole picture. The honest take is that GPT-5.5 wins on agentic terminal work and knowledge-work agents, ties or loses on some coding benchmarks, and dominates long-context retrieval at 1M tokens.

Benchmark	GPT-5.5	GPT-5.4	Opus 4.7	Gemini 3.1 Pro
Terminal-Bench 2.0	82.7%	75.1%	69.4%	68.5%
SWE-Bench Pro (Public)	58.6%	57.7%	64.3%	54.2%
Expert-SWE (OpenAI internal)	73.1%	68.5%	–	–
OSWorld-Verified	78.7%	75.0%	78.0%	–
GDPval (wins/ties)	84.9%	83.0%	80.3%	67.3%
Tau2-Bench Telecom	98.0%	–	–	–

Source: Vellum benchmark summary; the SWE-Bench Pro row is the public version with asterisk caveats from OpenAI’s own release.

Three observations worth flagging:

• Terminal-Bench 2.0 is a decisive lead for GPT-5.5. At 82.7% versus Opus 4.7’s 69.4%, the gap is 13+ points. For long-running terminal agents, pipeline runners, and DevOps automation, this benchmark is more representative of real agentic work than SWE-Bench Pro (Vellum).

• Opus 4.7 still wins SWE-Bench Pro nominally at 64.3% versus 58.6%, but the benchmark has known memorization caveats. Run your own eval on your own repos before committing (Alex Lavaee’s honest take).

• Long-context retrieval is where GPT-5.5 truly pulls ahead. On OpenAI MRCR v2 8-needle from 512K to 1M tokens, GPT-5.5 scores 74.0% versus GPT-5.4’s 36.6%. That is roughly a doubling of the prior generation’s accuracy at the same context length (Alex Lavaee, citing OpenAI announcement).

For the long-context angle in particular, our LLM Context Windows: Why Your 1M-Token Model Only Uses 32K piece covered the same retrieval cliff from a different angle earlier in 2026. The new MRCR numbers suggest the cliff is now meaningfully less steep for GPT-5.5, though still real.

Cost, latency, and when to spend the extra tokens

Pricing has shifted. Standard short-context GPT-5.5 is $5 input / $30 output per million tokens, with cached input at $0.50 (a 90% discount) and long-context prompts over 272K tokens priced at 2x input and 1.5x output (Morph pricing table). GPT-5.5 Pro is $30 input / $180 output – six times the standard tier.

The instinct for many practitioners will be to default to GPT-5.5 for everything because it is the new frontier. That is the wrong reflex. Three rules from production deployments:

1. Use Instant (GPT-5.3) for high-volume batch work where latency matters. At one fifth the price of GPT-5.5 for input tokens and the same per-token latency envelope as the prior generation for most tasks, it is the right tier for any task you would previously have called GPT-4o mini for. See our AI API Pricing in 2026 analysis for the broader pattern.

1. Set reasoning_effort deliberately, not lazily. For a simple look-and-respond prompt, low gives you the reasoning architecture without the token burn. For agentic multi-step work, high is the documented starting point. xhigh should only ship after you have run an eval showing it beats high on your specific task.

1. Cache aggressively. The 90% cached-input discount is the single biggest cost lever in the API. If you have a stable system prompt over 1K tokens, caching it is worth more than the entire token savings from switching models.

For the API cost trajectory, our AI Inference Cost in 2026 post tracks the year-over-year curve and the effect of caching on real workloads.

The cybersecurity angle: why GPT-5.5 carries a ‘High’ risk classification

This is the part of the launch that does not show up in the headline benchmark tables. OpenAI explicitly classifies GPT-5.5 as a “High” cyber risk – one tier below “Critical” – and shipped stricter classifiers as part of the rollout (CNBC, OpenAI announcement).

The reason is the competitive landscape. Anthropic publicly disclosed Claude Mythos Preview on April 7, 2026 with a restricted rollout to vetted cybersecurity partners under Project Glasswing (Wikipedia). Independent evaluation by the UK AI Security Institute found Mythos is the first AI model to complete a 32-step simulated end-to-end intrusion on a corporate network autonomously – chaining weaknesses, traversing the network, and reaching the defined compromise objective (GBHackers). The Mythos 5 / Fable 5 launch in June expanded access; by August, Mythos 5.1 had become the restricted-access cyber variant alongside the more permissive Claude Fable 5.1.

GPT-5.5 is OpenAI’s response to that landscape. The launch post notes that with GPT-5.2 in December 2025, OpenAI deployed cyber safeguards; with GPT-5.5, they are tightening those classifiers further. OpenAI Vice President of Research Mia Glaese confirmed that GPT-5.5 underwent “extensive third-party safeguard testing and red teaming for cyber and bio risks” (CNBC). For a deeper take on what GPT-5.5’s cyber posture means for the broader arms race, the Mr. Technology network has a payload on the topic: OpenAI Just Militarized GPT-5.5. Here’s Why That Matters..

The practical implication for non-cyber practitioners: do not be surprised if you occasionally hit stricter refusals on dual-use queries, especially anything that walks the line between defensive security analysis and exploit description. The model is the same; the classifiers are tighter.

Practical workflow: pick the right effort before you pick the model

The biggest shift in April 2026 is mental. The decision tree you used to navigate (“is this an o3 problem or a 4o problem?”) is gone. The replacement is shorter and more honest:

1. Is this latency-sensitive? Yes – GPT-5.3 Instant, reasoning effort low or none. No – keep reading.

1. Is the task multi-step or does it require tool use or planning? Yes – GPT-5.5 Thinking, reasoning effort high . No – GPT-5.4 Thinking at medium is often the better choice.

1. Have you tried high and still get wrong answers? Yes – move to GPT-5.5 Pro in ChatGPT, or gpt-5.5-pro in the API. Run your own eval before committing.

1. Are you on a tight token budget or running this at scale? Yes – cache aggressively and stay on GPT-5.4 Thinking.

That is it. Four questions. The old decision tree had you choose between named models for every distinct task class; the new tree asks you to pick effort and capability tier separately.

If you want a concrete starting point for the agentic pattern, our Building Production AI Agents: A Practical Guide to the OpenAI Agents SDK in 2026 and Complete Guide to AI Agents in 2026 walk through the patterns that pair naturally with the new effort controls.

Frequently Asked Questions

What happened to OpenAI’s o-series reasoning models?

OpenAI retired o4-mini on February 13, 2026 and o3 on May 28, 2026, with the August 26, 2026 sunset. The reasoning capability those models provided is now built into GPT-5.x Thinking variants – there is no separate o-series model selector in ChatGPT or the Responses API. For code that still references “o3” or “o4-mini” by name, switch to gpt-5.5 with reasoning: { effort: "high" } for the closest behavioral equivalent.

Did GPT-5.5 replace GPT-5.4?

GPT-5.5 launched April 23, 2026 as the new frontier tier and is available in ChatGPT (Plus/Pro/Business/Enterprise) and the API. GPT-5.4 Thinking remains available as the professional workhorse tier and is the default for users who do not actively select Thinking. The two coexist – GPT-5.4 Thinking is the right choice for tasks where you do not need GPT-5.5’s agentic coding or computer-use step up.

What is the difference between GPT-5.5 Thinking and GPT-5.5 Pro?

GPT-5.5 Thinking is the reasoning-capable tier for Plus, Pro, Business, and Enterprise users with adjustable thinking effort (Standard/Extended, plus Light/Heavy for Pro). GPT-5.5 Pro is the highest-capability variant for Pro/Business/Enterprise, using parallel test-time compute – the equivalent of the old o-series “pro” mode. Pro is six times the price per token of standard GPT-5.5.

How do I set reasoning effort in the API?

Use the Responses API with `reasoning: { effort: “low” | “medium” | “high” | “xhigh” }`. GPT-5.5 defaults to `medium`. Use `low` for latency-sensitive calls (voice, classification), `high` for complex agentic work, and `xhigh` only when your evals show it pays off – it materially increases latency and token cost. Always check the model page for current defaults; they are model-dependent.

How much does GPT-5.5 cost per million tokens?

Standard short-context: $5 input / $30 output per million tokens. Cached input is $0.50 (a 90% discount). Long-context prompts over 272K input tokens cost 2x input and 1.5x output. GPT-5.5 Pro is $30/$180 per million tokens – six times the standard tier. The Batch API halves both rates if latency is not a constraint.

Should I migrate production calls off o3 before the August 26 sunset?

Yes if you can. The behavioral mapping is not exact – GPT-5.5 Thinking with reasoning_effort: "high" is the closest analog for o3, and GPT-5.5 Pro with parallel test-time compute is the analog for o3-pro. Plan at least a week for eval and shadow traffic before the deadline. The default effort on GPT-5.5 is medium , not the highest tier, so you may also need to explicitly set high to match o3’s behavior on your specific tasks.

---

The model picker as described above reflects the April 28, 2026 rollout. OpenAI’s model lineup shifts quarterly; verify the current model picker and the per-model reasoning_effort defaults at platform.openai.com/docs/models before committing to architecture decisions.