GPT-5 Thinking: Enhanced Reasoning

• GPT-5 Thinking is an advanced reasoning mode in GPT-5 that dynamically allocates extra compute for step-by-step and multimodal synthesis in diverse challenges.
• It operates within a system-of-models architecture that routes queries based on difficulty, improving performance in clinical, planning, and spatial tasks.
• Empirical studies show GPT-5 Thinking excels in reasoning-intensive benchmarks but remains limited on specialized perception and long-context tasks.

GPT-5 Thinking denotes the higher-deliberation component of GPT-5 and, in recent research usage, a broader family of GPT-5 deployments that allocate extra test-time compute to step-by-step reasoning, multimodal synthesis, or hard problem solving. The literature describes it both as a deeper reasoning mode inside a system-of-models architecture and as an evaluation regime in which GPT-5 is explicitly prompted or routed to reason more extensively before committing to an answer. Across medical QA, planning, spatial benchmarks, mathematical experiments, and scientific case studies, the resulting empirical picture is strong but uneven: GPT-5 frequently leads peer models on reasoning-intensive tasks, yet it remains clearly limited on specialized perception-heavy subtasks, very long noisy contexts, and mathematically delicate cross-paper synthesis (Georgiou, 16 Aug 2025, Zhang et al., 18 Aug 2025, Wang et al., 11 Aug 2025).

1. Conceptualization and system architecture

In the papers surveyed here, GPT-5 Thinking is not treated as a single narrow feature. One line of work describes GPT-5 as a “unified system” composed of a smart, efficient model for most questions, a deeper reasoning model identified as “GPT-5 thinking” for harder problems, and a real-time router that selects between them at inference time. A related human-rated comparison characterizes GPT-5 more generally as a system-of-models architecture designed for task-specific optimization, with especially large gains in domains where context sensitivity, safety awareness, and tailored reasoning matter most (Zhang et al., 18 Aug 2025, Georgiou, 16 Aug 2025).

This framing places GPT-5 Thinking within a larger shift from monolithic inference toward conditional compute allocation. The central idea is that difficult queries should not be forced through the same inference pathway as easy ones. That idea later reappears in routing work, trace-retrieval work, and multi-round reasoning systems, all of which treat reasoning quality as partly a function of how compute is distributed across tasks, stages, or candidate solution paths.

A useful adjacent vocabulary comes from “synthetic deliberation,” which defines LLM-supported multi-perspective discourse as an externalized deliberative process that separates compartmentalization from integration. In that formalism, distinct agents explore different regions of a rugged solution landscape, while a human or controlling layer integrates them, with the interaction parameter $a$ or $\alpha$ controlling the balance between independence and convergence. Although this is not a paper about GPT-5 specifically, it supplies a theoretical frame for why additional reasoning budget, preserved viewpoint distinctiveness, and controlled synthesis may improve hard problem solving (Park et al., 4 Jan 2025).

2. Operational forms and evaluation protocols

Much of the GPT-5 Thinking literature is methodologically explicit about isolating capability rather than prompt engineering. In the principal medical zero-shot chain-of-thought setup, every sample begins with the system prompt “You are a helpful medical assistant.” The first user turn contains the question and the phrase “Let’s think step by step.” GPT then emits a free-form rationale, often stored as prediction_rationale. A second user turn constrains the answer to a final option letter: “Therefore, among A through {END_LETTER}, the answer is,” after which only the discrete answer is scored. For multimodal items, images are inserted into the first user turn as image_url entries, so reasoning proceeds jointly over text and images while evaluation uses only the constrained final answer. This protocol is used across MedQA, MedXpertQA, MMLU medical subsets, USMLE self-assessment exams, VQA-RAD, and a BraTS-derived brain-tumor MRI benchmark (Wang et al., 11 Aug 2025, Safari et al., 14 Aug 2025).

Other evaluations use different operationalizations. A radiology and radiation-oncology benchmark explicitly removes chain-of-thought and instead tests zero-shot, single-turn interactions with no fine-tuning, no in-context exemplars, no retrieval augmentation, and no step-by-step reasoning chain. By contrast, end-to-end planning from PDDL uses few-shot prompting with general instructions, domain and task files, a checklist of common pitfalls, two illustrative examples from Gripper and Logistics, and external validation with VAL. Spatial-intelligence work standardizes prompts, answer formatting, and extraction across eight benchmarks, using zero-shot CoT plus a three-stage extraction pipeline that falls back from rule-based parsing to LLM-assisted extraction when needed (Hu et al., 15 Aug 2025, Corrêa et al., 12 Nov 2025, Cai et al., 18 Aug 2025).

These protocol differences matter because “GPT-5 Thinking” in the literature can denote either a model-side reasoning mode or a benchmark-side evaluation choice. Some studies therefore test the gains from explicit rationalization; others test GPT-5 as a stronger general multimodal model even when reasoning traces are suppressed at output time.

3. Multimodal clinical reasoning

The most sustained empirical case for GPT-5 Thinking appears in medicine. A large multimodal medical-reasoning study evaluates GPT-5, GPT-5-mini, GPT-5-nano, and GPT-4o-2024-11-20 under a single standardized zero-shot CoT protocol across MedQA, MedXpertQA Text and MM, MMLU-Medical, USMLE self-assessment exams, and VQA-RAD. GPT-5 reaches 95.84% on MedQA versus GPT-4o’s 91.04%; on MedXpertQA Text it scores 56.96% on Reasoning and 54.84% on Understanding versus 30.63% and 29.54%; on USMLE it reaches 93.28% on Step 1, 97.50% on Step 2, 94.89% on Step 3, and 95.22% overall average. The strongest gains occur on MedXpertQA MM, where GPT-5 attains 69.99% Reasoning and 74.37% Understanding versus GPT-4o’s 40.73% and 48.19%. The paper notes that a summary request mentions larger abstract-level gains, but its detailed table and narrative consistently report the safer values of +29.26 percentage points for Reasoning and +26.18 for Understanding. The same study also compares against pre-licensed human experts: on MedXpertQA MM, GPT-5 exceeds the human baseline by +24.23% in Reasoning and +29.40% in Understanding, and a representative case study shows coherent synthesis of vomiting history, epigastric tenderness, elevated lipase, imaging findings, suprasternal crepitus, and blood-streaked emesis into a diagnosis of esophageal perforation with recommendation of a Gastrografin swallow (Wang et al., 11 Aug 2025).

A later landscape commentary broadens that picture across 14 clinically grounded datasets. It reiterates the large MedXpertQA gains, reports modest but consistent USMLE and MedQA improvements, and then shows that multimodal medical reasoning is heterogeneous rather than uniformly solved. GPT-5 is competitive or state-of-the-art on several VQA tasks, and on mammography it improves over GPT-4o by margins ranging from roughly 10 to 40 percentage points on several EMBED, InBreast, and CBIS-DDSM subtasks. Yet the same commentary emphasizes that specialized mammography systems remain far stronger, citing examples such as EMBED malignancy at 82.3% for Mammo-CLIP versus 52.8% for GPT-5, and InBreast BI-RADS at 90.6% for a multi-scale region selection network versus 36.9% for GPT-5. The paper’s core conclusion is therefore dual: GPT-5 advances integrated multimodal clinical reasoning, but generalist zero-shot systems do not replace purpose-built perception models in highly specialized settings (Florea et al., 5 Mar 2026).

A targeted evaluation in radiology, radiation oncology, and medical physics reaches a similar conclusion from a narrower benchmark design. There GPT-5 is best across VQA-RAD, SLAKE, and an in-house 150-question Medical Physics Board Examination-style set. The reported scores are 74.90% on VQA-RAD, 88.60% aggregate on SLAKE with 86.20% on open-ended and 92.31% on closed-ended items, and 136/150 or 90.7% on the board-style physics questions. The paper highlights especially large gains over GPT-4o in chest-mediastinal (+20.00%), lung-focused (+13.60%), and brain-tissue (+11.44%) questions, and notes that 90.7% exceeds the estimated ABR Part I medical physics passing threshold of roughly 70% to 75% (Hu et al., 15 Aug 2025).

At the same time, not all clinical multimodal benchmarks validate an “above-expert” narrative. In brain-tumor MRI reasoning on BraTS-derived cohorts, GPT-5 reaches 42.68% on MET, 46.34% on GLI, 42.12% on MEN, and 43.71% macro-average. GPT-5-mini slightly exceeds it overall at 44.19%, GPT-4o is close behind at 41.49%, and the authors judge all of these scores moderate rather than clinically adequate. They explicitly state that the GPT-5 family can handle some structured neuro-oncology reasoning, but not at a level acceptable for clinical use (Safari et al., 14 Aug 2025).

4. Performance across general reasoning benchmarks

Outside medicine, GPT-5 Thinking is evaluated in both expert-rated and benchmarked settings. In a human-rated study covering lesson planning, assignment evaluation, clinical diagnosis, research generation, and ethical reasoning, mixed-effects models show that GPT-5 significantly outperforms GPT-4 in lesson planning, clinical diagnosis, research generation, and ethical reasoning, while assignment evaluation shows no significant overall difference. The strongest gains appear in clinical diagnosis with $B = 1.17$, $SE = 0.34$, $t = 3.46$, $p < 0.01$, and in research generation with $B = 1.07$, $SE = 0.22$, $t = 4.91$, $p < 0.001$. By contrast, assignment evaluation yields $\alpha$0, $\alpha$1, $\alpha$2, $\alpha$3, which the paper interprets as a plateau domain rather than a decisive GPT-5 advantage (Georgiou, 16 Aug 2025).

Structured reasoning benchmarks reveal similar strengths and ceilings. In end-to-end planning from PDDL, GPT-5 generates plans directly from domain and task descriptions, and those plans are checked with VAL. On 360 tasks from eight IPC 2023 Learning Track domains, GPT-5 solves 205 standard tasks, essentially matching LAMA’s 204 and exceeding DeepSeek R1’s 157 and Gemini 2.5 Pro’s 155. Under obfuscation, where all actions, predicates, and objects are renamed with random strings, GPT-5 drops to 152 solved tasks, still the best LLM but clearly below LAMA’s unchanged 204. The same study therefore portrays GPT-5 as roughly planner-competitive in aggregate coverage on standard tasks, yet still less robust to symbol-semantic removal than a classical planner (Corrêa et al., 12 Nov 2025).

Spatial-intelligence studies are more cautionary. One large evaluation across eight benchmarks and more than one billion total tokens identifies GPT-5 as the strongest spatial model tested, with unified chance-adjusted accuracy scores including 36.27 on VSI-Bench, 64.18 on SITE, 22.47 on MMSI, 50.24 on OmniSpatial, and 15.96 on SpatialViz. Yet representative human comparisons remain wide: 53.14 versus 96.58 on VSI-Bench, 62.70 versus 92.3 on OmniSpatial, 56.67 versus 96.50 on STARE, and 36.97 versus 82.46 on SpatialViz. The taxonomy proposed in that paper locates the major residual deficits in mental reconstruction, perspective-taking, deformation and assembly, and comprehensive reasoning. The authors’ broader claim is that GPT-5 has achieved unprecedented strength in spatial intelligence, but not human-level spatial intelligence (Cai et al., 18 Aug 2025).

Zero-shot object counting offers a narrower but revealing case. GPT-5, driven only by textual prompts plus the full image and returning a plain integer count, is often competitive with state-of-the-art zero-shot counting approaches on FSC-147 and usually more accurate than GPT-4o in the more stable prompt settings. On CARPK, however, both GPT models remain much worse than counting-specific systems such as RCC, and in extremely crowded scenes both models can fail catastrophically: the paper reports lego pieces with ground truth 2560 for which both GPT-4o and GPT-5 predict 1 (Füzesséry et al., 2 Dec 2025).

Educational assessment work also shows that GPT-5’s advantages are sensitive to label structure. In automated scoring of critical-thinking subskills in argumentative essays, GPT-5 with few-shot prompting achieves the strongest results, with accuracy 0.554, RMSE 0.990, macro F1 0.476, weighted F1 0.554, and Krippendorff’s $\alpha$4. That is better than GPT-5 zero-shot, GPT-5-mini, and ModernBERT, and it is the only system that reaches the paper’s exploratory reliability target. The gains are concentrated in subskills with frequent and separable categories; rare or subtle categories, especially Exemplifying, remain difficult (Peczuh et al., 14 Oct 2025).

Long-context evaluation reveals a different failure mode altogether. On a depression-detection task built from 20K Twitter posts, model performance begins to degrade once the input exceeds 5K tweets, estimated at more than 70K tokens, and at 20K tweets or roughly 300K tokens all models fall to about 50–53% accuracy. GPT-5’s distinctive property in that setting is not preserved accuracy but preserved precision, which remains around 95%. The paper treats that as potentially useful for sensitive applications, while emphasizing the broader gap between nominal context-window capacity and robust long-context utilization (Esmi et al., 15 Feb 2026).

5. Mathematical and scientific reasoning

Research-oriented evaluations probe whether GPT-5 Thinking extends beyond benchmark-solving into literature-conditioned reasoning. The proposed “Gödel Test” asks whether a model can prove very simple, genuinely novel conjectures in combinatorial optimization after reading one or two source papers and receiving minimal hinting. On five such problems, GPT-5 produces nearly correct solutions on the three easier cases; on Problem 2 it derives a different approximation guarantee that, after manual checking, refutes the authors’ original conjecture while still yielding a valid result; on Problem 4 it fails; and on Problem 5 it proposes the same algorithm the authors had in mind but fails in the analysis. The authors interpret this as evidence of meaningful progress on routine mathematical reasoning and occasional originality, with a clear breakdown when cross-paper synthesis becomes essential (Feldman et al., 22 Sep 2025).

A controlled Malliavin–Stein experiment reaches a related but more skeptical conclusion. There the target was to transform a qualitative fourth-moment theorem into a quantitative total-variation bound with an explicit rate depending on the fourth cumulant, first in the Gaussian setting and then in the Poisson setting. GPT-5 quickly produced a largely correct strategy for the Gaussian case but made a crucial error in the mixed contraction comparison; it corrected course only after the error was challenged directly. In the Poisson case, it noticed that mixed odd moments need not vanish automatically, but initially missed the positivity fact needed to complete the argument and incorporated it only after being pointed to the relevant equation. The authors therefore present GPT-5 as a useful incremental research assistant and drafting partner, but not as an autonomous guarantor of proof correctness (Diez et al., 3 Sep 2025).

A broader collection of scientific case studies pushes the same interpretation across disciplines. In mathematics, physics, astronomy, computer science, biology, and materials science, GPT-5 is reported to have produced new, concrete steps in ongoing research, with four new mathematical results carefully verified by the human authors. The paper emphasizes that GPT-5 contributed literature search, proof ideas, derivation steps, experimental next actions, and candidate hypotheses, while human experts remained essential for selecting problems, validating arguments, checking novelty, and determining final scientific significance (Bubeck et al., 20 Nov 2025).

Taken together, these studies distinguish between two forms of research competence. GPT-5 is often effective at extension, adaptation, recombination, and acceleration. It is much less reliable when novelty depends on delicate invariant-tracking, hidden side conditions, or synthesis across multiple proof templates.

6. Retrieval, routing, parallel thinking, and unresolved issues

A large secondary literature studies how GPT-5-style reasoning can be improved by changing the structure around the model rather than the base model alone. Routing is the clearest example. Avengers-Pro begins from the same premise attributed to GPT-5: not every query should be handled by the same model. It embeds each query with Qwen3-embedding-8B, clusters the space into $\alpha$5 semantic groups, estimates per-cluster performance and cost profiles, and routes with the score

$\alpha$6

In the reported operating points, Avengers-Pro reaches 66.66% average accuracy at roughly GPT-5-medium cost, compared with GPT-5-medium’s 62.25%, or matches that baseline’s accuracy at 27% lower cost. The paper’s significance for GPT-5 Thinking is conceptual as much as empirical: it treats “thinking” as an inference-time allocation problem rather than a single fixed model mode (Zhang et al., 18 Aug 2025).

Retrieval-augmented reasoning work makes a parallel argument about corpus choice. Instead of retrieving web documents, the T3 framework retrieves prior thinking traces. For GPT-5 on AIME, the No RAG baseline is 86.7, retrieval over raw chunked Gemini-2-thinking traces raises this to 91.7, and T3-Reflect reaches 93.3, which the paper describes as a +7.6% relative gain over No RAG. The same study reports a cost-accuracy shift for GPT-5 from 76.14 to 80.53 while reducing inference cost from 1.22 to 1.04 cents per query. The implication is that retrieving “how to reason” can improve frontier reasoning even when the solver is already strong (Arabzadeh et al., 5 May 2026).

Parallel-thinking systems generalize the same principle from retrieval to search. A competitive-programming study argues that reasoning tokens can be scaled not only by making one chain longer, but by distributing them across many threads and rounds of generation, verification, and refinement. Its full system, starting from Seed-OSS-36B and using 16 threads and 16 rounds per thread, matches the underlying RL model’s oracle pass@16 at pass@1 with 7.6 million tokens per problem on average and surpasses GPT-5-high on 456 hard AetherCode problems. Although this is not a GPT-5 system paper, it is directly relevant to the meaning of “thinking”: it shows that stronger reasoning may emerge from structured parallelism rather than from a single monolithic rationale (Zhang et al., 1 Apr 2026).

The unresolved issues are strikingly consistent across domains. Clinical papers call for prospective clinical trials, calibration methods, and domain-adapted fine-tuning before any safe deployment; spatial studies show large residual deficits in perspective-taking and deformation reasoning; long-context work demonstrates sharp degradation well before nominal context limits; and benchmark success frequently coexists with hallucination risk, unstable small-dataset calibration, or absent uncertainty estimates. GPT-5 Thinking therefore names a real and measurable advance in test-time reasoning, but not a finished solution to robustness, transparency, or domain-specific reliability (Wang et al., 11 Aug 2025, Cai et al., 18 Aug 2025, Esmi et al., 15 Feb 2026).