Passer au contenu principal
OpenAI

16 avril 2025

VersionsProduit

Introducing OpenAI o3 and o4-mini

Chargement…

Mise à jour le 10 juin 2025 : OpenAI o3‑pro est maintenant disponible pour les utilisateurs Pro dans ChatGPT, ainsi que dans notre API. Comme OpenAI o1‑pro, o3‑pro est une version de notre modèle le plus intelligent, OpenAI o3, conçu pour réfléchir plus longtemps et fournir les réponses les plus fiables. Vous trouverez tous les détails dans nos notes de version(s'ouvre dans une nouvelle fenêtre).


Aujourd'hui, nous lançons OpenAI o3 et o4-mini, les derniers modèles de notre série o, conçus pour réfléchir plus longtemps avant de répondre. Ce sont les modèles les plus avancés que nous ayons publiés à ce jour, marquant une étape importante dans les capacités de ChatGPT, aussi bien pour les utilisateurs curieux que pour les chercheurs expérimentés. Pour la première fois, nos modèles de raisonnement peuvent utiliser et combiner de manière autonome l'ensemble des outils de ChatGPT : recherche sur le Web, analyse de fichiers importés et de données avec Python, raisonnement approfondi sur les éléments visuels et génération d'images. Ces modèles sont entraînés à déterminer quand et comment utiliser les outils pour produire des réponses précises et réfléchies, dans le format approprié, généralement en moins d'une minute, pour résoudre des problèmes complexes. Cela leur permet de traiter plus efficacement des questions à plusieurs volets, une étape vers un ChatGPT plus agentique, capable d'exécuter certaines tâches de manière autonome. La combinaison d'un raisonnement de pointe et d'un accès complet aux outils se traduit par des performances nettement supérieures dans les tests académiques et les cas d'usage concrets, établissant une nouvelle référence en matière d'intelligence et d'utilité.

Ce qui a changé

OpenAI o3 est notre modèle de raisonnement le plus puissant, qui repousse les frontières dans les domaines du codage, des mathématiques, des sciences, de la perception visuelle et bien plus encore. Il constitue ainsi une nouvelle référence sur diverses évaluations, notamment Codeforces, SWE-bench (sans infrastructure spécifique) et MMMU. Il offre une solution idéale pour les demandes complexes imposant une analyse sur plusieurs niveaux et dont les réponses ne sont pas forcément évidentes. Il se montre particulièrement efficace sur les tâches visuelles, comme l’analyse d’images, de tableaux et de graphiques. Dans les évaluations menées par des experts externes, o3 commet 20 % d’erreurs majeures en moins qu’OpenAI o1 sur des tâches difficiles et concrètes — excellant particulièrement dans des domaines comme la programmation, le conseil aux entreprises et la recherche d’idées créatives. Ses premiers testeurs ont mis en avant la rigueur de ses analyses dans le cadre d’échanges d’idées et sa capacité à générer et évaluer de manière critique des hypothèses inédites, notamment dans les domaines de la biologie, des mathématiques et de l’ingénierie.

OpenAI o4-mini est un modèle plus petit, optimisé pour un raisonnement rapide et économique — il offre des performances remarquables pour sa taille et son coût, notamment en mathématiques, programmation et tâches visuelles. Il s’agit du modèle le plus performant évalué lors des éditions 2024 et 2025 de l’AIME. Bien que l’accès à un ordinateur réduise de manière non négligeable la difficulté de l’examen AIME, nous avons également trouvé remarquable qu’o4-mini atteigne un taux de réussite de 99,5 % pass@1 (100 % consensus@8) sur l’édition 2025 de l’AIME lorsqu’il a accès à un interpréteur Python. Bien que ces résultats ne doivent pas être comparés aux performances des modèles sans accès aux outils, ils illustrent l'efficacité avec laquelle o4-mini exploite les outils disponibles ; o3 affiche des améliorations similaires sur AIME 2025 grâce à l'utilisation d'outils (98,4 % de réussite@1, 100 % de consensus@8).

Lors d’évaluations réalisées par des experts, o4-mini s’est aussi montré plus performant que son prédécesseur (o3‑mini) sur des tâches hors STIM et des domaines comme la science des données. Grâce à son efficience, o4-mini permet des limites d’utilisation nettement plus élevées que o3, ce qui en fait une option solide pour un volume élevé et un débit important, adaptée aux questions qui bénéficient d’un raisonnement. Des experts externes ont évalué les deux modèles comme démontrant une amélioration du suivi des instructions et des réponses plus utiles et vérifiables que leurs prédécesseurs, grâce à une intelligence améliorée et à l’inclusion de sources Web. Par rapport aux versions précédentes de nos modèles de raisonnement, ces deux modèles se montrent aussi plus naturels et conversationnels, notamment car ils font appel à des souvenirs et des conversations passées pour rendre leurs réponses plus personnalisées et pertinentes.

Multimodal

Codage

Toutes les évaluations SWE-bench utilisent un sous-ensemble fixe de n=477 tâches vérifiées qui ont été validées sur notre infrastructure interne.

Respect des instructions et utilisation agentique d’outils

Tous les modèles sont évalués sur des paramètres d’« effort de raisonnement » élevé — similaire à des variantes comme « o4-mini-high » dans ChatGPT.

Continuing to scale reinforcement learning

Throughout the development of OpenAI o3, we’ve observed that large-scale reinforcement learning exhibits the same “more compute = better performance” trend observed in GPT‑series pretraining. By retracing the scaling path—this time in RL—we’ve pushed an additional order of magnitude in both training compute and inference-time reasoning, yet still see clear performance gains, validating that the models’ performance continues to improve the more they’re allowed to think. At equal latency and cost with OpenAI o1, o3 delivers higher performance in ChatGPT—and we've validated that if we let it think longer, its performance keeps climbing.

We also trained both models to use tools through reinforcement learning—teaching them not just how to use tools, but to reason about when to use them. Their ability to deploy tools based on desired outcomes makes them more capable in open-ended situations—particularly those involving visual reasoning and multi-step workflows. This improvement is reflected both in academic benchmarks and real-world tasks, as reported by early testers.

Réfléchir avec des images

ChatGPT thinking with images

Pour la première fois, ces modèles peuvent intégrer des images directement dans leur chaîne de réflexion. Ils ne se contentent pas de voir une image, ils l’intègrent dans leur réflexion. Cela ouvre la voie à une nouvelle catégorie de résolution de problèmes qui combine le raisonnement visuel et textuel, comme en témoignent leurs performances de pointe dans les références multimodales.

Les gens peuvent téléverser une photo d’un tableau blanc, un schéma de manuel ou un croquis fait à la main, et le modèle peut l’interpréter — même si l’image est floue, inversée ou de mauvaise qualité. À l’aide d’outils, les modèles peuvent manipuler les images à la volée — les faire pivoter, zoomer ou transformer dans le cadre de leur processus de raisonnement.

Ces modèles offrent une précision exemplaire sur les tâches de perception visuelle, ce qui leur permet de répondre à des questions hors de leur portée jusqu’à présent. Consultez le blogue de recherche sur le raisonnement visuel pour en savoir plus.

Vers l’utilisation agentique d’outils

OpenAI o3 et o4-mini disposent d’un accès complet aux outils intégrés dans ChatGPT, ainsi qu’à vos propres outils personnalisés via l’appel de fonctions dans l’API. Ces modèles sont formés à réfléchir à la manière de résoudre des problèmes, en choisissant quand et comment utiliser des outils pour produire rapidement des réponses détaillées et réfléchies dans les bons formats de sortie — généralement en moins d’une minute.

Par exemple, un utilisateur pourrait demander : « Comment la consommation d'énergie estivale en Californie se comparera-t-elle à celle de l'année dernière? » Le modèle peut chercher sur le Web des données des services publics, écrire un code Python pour établir une prévision, générer un graphique ou une image, et expliquer les principaux facteurs derrière la prédiction, en enchaînant plusieurs appels à des outils. Le raisonnement permet aux modèles de réagir et de changer de direction si les informations qu’ils trouvent le justifient. Par exemple, ils peuvent effectuer plusieurs recherches Web à l’aide de moteurs de recherche, consulter les résultats et essayer de nouvelles recherches s’ils ont besoin de plus d’informations.

Cette approche flexible et stratégique permet aux modèles de gérer des tâches qui demandent un accès à des informations plus récentes que celles dont ils disposent, un raisonnement étendu, la synthèse d’informations et la génération de sorties dans diverses modalités.

Tous les exemples ci-dessous ont été réalisés avec OpenAI o3.

OpenAI o3

Solve the following very hard math problem using your in-built knowledge of the literature:

Construct a degree 19 polynomial \ (p(x)) \ in \ mathbb{C}[x] such that \ (X := \ {\, p(x) = p(y) \, \} \ subset \ mathbb{P}^{1} \ times \ mathbb{P}^{1}\ ) has at least 3 (but not all linear) irreducible components over \ mathbb{C}. Choose \ (p(x)\ ) to be odd, monic, have real coefficients and linear coefficient \ (-19\ ) and calculate \ (p(19)\ ).

Raisonné sur la construction polynomiale pendant 55 secondes
  • The curve defined by p(x) = p(y) factors if p(x) is decomposable or has repeated factors. Actually, p(x)-p(y) always factors, and it’s divisible by (x-y). This gives one factor, and the rest is a degree d-1 polynomial in x and y. For the curve to have at least three irreducible components, p(x)-p(y) must factor into at least three nontrivial factors. Each factor corresponds to a component in projective space, and at least one must be nonlinear.
  • I’m checking the Dickson polynomials D_n(x,a) and focusing on their coefficients. With a=1, it seems the leading coefficient is always 1, confirming that Dickson polynomials are monic. The linear coefficient for odd n is following a predictable sign pattern: for n=19, it gives a coefficient of -19, which is exactly what I need. So the Dickson polynomial D_n(x,1) with odd n fits nicely with the linear coefficient I was looking for.
  • I've confirmed that the linear coefficient for p(x)=D19(x)p(x) = D_{19}(x)p(x)=D19​(x) is correct, with the necessary scaling factor to match. Now for the next step, I need to compute D19(19)D_{19}(19)D19​(19). Since Tn(z)T_n(z)Tn​(z) grows quickly for real z>1z > 1z>1, I plan to compute this value using Python, leveraging the recurrence to avoid mistakes—it's going to be a big number. I'll make sure to present the numeric result neatly so it's understandable since the coefficients are integers.
  • The Dickson polynomial D_{19}(x,1) is confirmed to have the required properties: monic, odd, real coefficients, and a linear coefficient of -19. The factorization of D_{19}(x) into irreducible components is also verified and satisfies the condition of having at least 10 components (1 linear + 9 quadratics). To compute p(19), we'll use the recurrence relation for Dickson polynomials, but there's also a closed-form expression involving cosine that could be handy here. I'll proceed accordingly!
  • I confirmed the linear coefficient is working correctly, which means now I need to calculate p(19)=D19(19)p(19) = D_{19}(19)p(19)=D19​(19). This involves evaluating T19(9.5)T_{19}(9.5)T19​(9.5), which grows rapidly. To ensure I get an accurate result, I'll compute this using recurrence in Python. The result is a large number: 1,876,572,071,974,094,803,391,179. I'll use the commas to make it easier to read: p(19)=1,876,572,071,974,094,803,391,179p(19) = 1,876,572,071,974,094,803,391,179p(19)=1,876,572,071,974,094,803,391,179.
  • For a = 1, there’s a relationship between Dickson polynomials and Chebyshev polynomials: D_n(x, 1) = 2^(n-1) T_n(x/2). Checking this, we see that the definition of D_n(u + 1/u) aligns well with Chebyshev’s formula. By setting x = 2 cos(θ), we find that D_n(x) equals 2 T_n(x/2). Importantly, the leading coefficient works out to 1, confirming the identity holds. This is nice!
  • Now, I need to verify that p(x)−p(y)p(x) - p(y)p(x)−p(y) factors as desired, relying on the known Dickson polynomial factorization. According to the literature (like Fried and Jarden), the difference Dn(x)−Dn(y)D_n(x) - D_n(y)Dn​(x)−Dn​(y) factors into a product involving (x−y)(x - y)(x−y) and other expressions. I'll provide the explicit polynomial D19(x,1)D_{19}(x, 1)D19​(x,1), list the coefficients, and explain how it satisfies the desired properties: monic, odd, etc. Lastly, I'll compute p(19)p(19)p(19).

Analyzed

Python

1
from math import comb, cos, acos
2
# Let's compute Dickson D_n(x,a=1) recursively
3
def dickson(n,x,a=1):
4
if n==0:
5
return 2
6
if n==1:
7
return x
8
Dm2=2
9
Dm1=x
10
for k in range(2,n+1):
11
D= x*Dm1 - a*Dm2
12
Dm2, Dm1 = Dm1, D
13
return Dm1
14

15
d19 = dickson(19,19,1)
16
d19

1876572071974094803391179

Analyzed

Python

1
d19

1876572071974094803391179

Analyzed

Python

1
import locale, sys, math, json, textwrap, re, decimal

Analyzed

Python

1
"{:,}".format(d19)

'1,876,572,071,974,094,803,391,179'

Analyzed

Python

1
# compute coefficients of Dickson polynomial D_n(x,1) via recurrence polynomial representation
2
from sympy import symbols, Poly
3

4
x = symbols('x')
5
def dickson_poly(n):

Poly(x**19 - 19*x**17 + 152*x**15 - 665*x**13 + 1 729*x**11 - 2 717*x**9 + 2 508*x**7 - 1 254*x**5 + 285*x**3 - 19*x, x,

Analyzed

Python

1
poly19.as_expr()

x**19 - 19*x**17 + 152*x**15 - 665*x**13 + 1 729*x**11 - 2 717*x**9 + 2 508*x**7 - 1 254*x**5 + 285*x**3 - 19*x

Done

A convenient way to force many factors in
p(x)    p(y)p(x) \;-\; p(y) is to use a Dickson polynomial of the first kind.
For a parameter
aCa \in \mathbb{C} in math these are defined by

D0 ⁣(x,a)=2,  D1 ⁣(x,a)=x,  Dn ⁣(x,a)=xDn1 ⁣(x,a)aDn2 ⁣(x,a)(n2),D_0\!\left(x,\, a\right) = 2,\; D_1\!\left(x,\, a\right) = x,\; D_n\!\left(x,\, a\right) = x\,D_{n-1}\!\left(x,\, a\right) - a\,D_{n-2}\!\left(x,\, a\right)\quad (n \ge 2),

and satisfy the functional identity

Dn ⁣(u+au,a)=un+anun.D_n\!\left(u + \frac{a}{u},\, a\right) = u^{n} + \frac{a^{n}}{u^{n}}.


1. Choosing the polynomial

Take a=1a = 1 and n=19n=19.
Because
nn is odd, D19 ⁣(x,1)D_{19}\!\left(x,\, 1\right) is an odd polynomial; the recurrence immediately shows it is monic and has integer (hence real) coefficients.
The coefficient of
xx is x=(1)n12n=19,x = (-1)^{\frac{n-1}{2}}\,n = -19, so all requested normalisations are satisfied.

Writing it out gives

p(x)=D19 ⁣(x,1)=x1919x17+152x15665x13+1729x112717x9+2508x71254x5+285x319x. \boxed{% p(x) = D_{19}\!\left(x,1\right) = x^{19} - 19x^{17} + 152x^{15} - 665x^{13} + 1729x^{11} - 2717x^{9} + 2508x^{7} - 1254x^{5} + 285x^{3} - 19x% }\,.


2. Irreducible components of fX  =  {p(x)=p(y)}fX \;=\; \bigl\{\, p(x) = p(y) \,\bigr\}

For Dickson polynomials one has the classical factorisation

Dn ⁣(x,1)    Dn ⁣(y,1)  =  (xy)j=1n12(x2+y22cos ⁣(2πjn)xy4sin2 ⁣(πjn)),D_n\!\left(x,1\right) \;-\; D_n\!\left(y,1\right) \;=\; (x - y)\, \prod_{j=1}^{\frac{n-1}{2}} \Bigl( x^{2} + y^{2} - 2\cos\!\bigl(\tfrac{2\pi j}{n}\bigr)\,xy - 4\sin^{2}\!\bigl(\tfrac{\pi j}{n}\bigr) \Bigr),

where each quadratic factor is irreducible over C\mathbb{C} (see, e.g., Lidl–Niederreiter, Finite Fields, §3.6).
For
n=19n=19 this gives

  • one linear component x=y;x=y;
  • nine quadratic components (corresponding to (j=1,,9).\bigl(j = 1,\,\ldots,\,9\bigr).

Hence XP1×P1X \subset \mathbb{P}^{1} \times \mathbb{P}^{1} has 1010 irreducible components, at least three of them and not all linear, as required.


3. Evaluating p(19)p(19)

Using the recurrence (or the identity Dn ⁣(x,1)  =  2Tn ⁣(x2)D_n\!\left(x,\,1\right) \;=\; 2\,T_n\!\bigl(\tfrac{x}{2}\bigr) with Chebyshev TnT_n one obtains

p(19)  =  D19 ⁣(19,1)  =  1876572071974094803391179.p(19) \;=\; D_{19}\!\left(19,\,1\right) \;=\; 1\,876\,572\,071\,974\,094\,803\,391\,179.


Thus the polynomial above fulfils every condition, and

p(19)  =  1,876,572,071,974,094,803,391,179.\boxed{% p(19) \;=\; 1{,}876{,}572{,}071{,}974{,}094{,}803{,}391{,}179 }.

OpenAI o1

Solve the following very hard math problem using your in-built knowledge of the literature:

Construct a degree 19 polynomial \ (p(x)) \ in \ mathbb{C}[x] such that \ (X := \ {\, p(x) = p(y) \, \} \ subset \ mathbb{P}^{1} \ times \ mathbb{P}^{1}\ ) has at least 3 (but not all linear) irreducible components over \ mathbb{C}. Choose \ (p(x)\ ) to be odd, monic, have real coefficients and linear coefficient \ (-19\ ) and calculate \ (p(19)\ ).

Réflexion durant 1 min et 19 s

I’m piecing together a degree 19 polynomial with only odd powers and a linear coefficient of -19, like: p(x)=x19+a17x17+…+a1xp(x) = x^{19} + a_{17}x^{17} + \ldots + a_1xp(x)=x19+a17​x17+…+a1​x.

I'm examining if the polynomial p(x)=x19+a17x17+…+a3x3−19xp(x) = x^{19} + a_{17}x^{17} + \ldots + a_3x^3 - 19xp(x)=x19+a17​x17+…+a3​x3−19x has a notable classical form, considering symmetry or factorization for multiple irreducible components.

To ensure p(x)p(x)p(x) has three or more irreducible components, I’m considering the factorization of p(x)−p(y)p(x) - p(y)p(x)−p(y) into at least three irreducible factors in C[x,y]\mathbb{C}[x,y]C[x,y].

I'm working through x19−y19x^{19} - y^{19}x19−y19 to ensure it balances linear and nonlinear factors, aiming for a polynomial with at least three irreducible components that aren't all linear.

Taking a closer look at how p(x)−p(y)p(x) - p(y)p(x)−p(y) can be factored into three distinct polynomials in xxx, aiming to achieve the desired irreducible components.

I'm thinking about how composing ppp with simpler polynomials, like xm−cx^m - cxm−c, could result in multiple irreducible components, achieving at least three distinct factors.

I’m exploring the factorization of p(x)−p(y)p(x) - p(y)p(x)−p(y) with cyclotomic polynomials. This method generates nnn linear irreducible factors, which might not achieve the required three distinct components.

I'm focusing on constructing an "odd" and monic polynomial, ensuring all exponents are odd and the leading term is x19x^{19}x19. This approach aims to achieve the desired factorization.

I'm working with the Chebyshev polynomials to ensure p(x)−p(y)p(x) - p(y)p(x)−p(y) has at least three irreducible components, leveraging their even or odd properties to manage coefficients.

Considering the odd Chebyshev polynomial T19(x)T_{19}(x)T19​(x) for potential factorization, while also examining composite polynomials of lower degrees to enhance factorization properties.

I'm analyzing how a polynomial composed of lower-degree polynomials, like (x3−3x)k(x^3 - 3x)^k(x3−3x)k, could yield p(x)−p(y)p(x) - p(y)p(x)−p(y) with at least three irreducible components.

Piecing together polynomials with integer coefficients, like (x3−3x)k(x^3 - 3x)^k(x3−3x)k, to achieve factorization in p(x)−p(y)p(x) - p(y)p(x)−p(y) with at least three irreducible components, aiming for degree 19.

I’m examining how p(x)=(x3−3x)kp(x) = (x^3 - 3x)^kp(x)=(x3−3x)k or p(x)=x19−19xp(x) = x^19 - 19xp(x)=x19−19x might yield p(x)−p(y)p(x) - p(y)p(x)−p(y) with three irreducible components, focusing on ensuring the polynomial is of degree 19.

I’m analyzing p(x)−p(y)p(x) - p(y)p(x)−p(y) into (x−y)(x18+x17y+…+y18−19)(x-y)(x^{18} + x^{17}y + \ldots + y^{18} - 19)(x−y)(x18+x17y+…+y18−19) to identify irreducible components. The polynomial factor x18+x17y+…+y18−19x^{18} + x^{17}y + \ldots + y^{18} - 19x18+x17y+…+y18−19 remains challenging to factor further.

I’m thinking through a new approach: adjusting the Chebyshev polynomial T19(x)T_{19}(x)T19​(x) to make it monic, aiming to achieve at least three irreducible components in p(x)−p(y)p(x) - p(y)p(x)−p(y).

OK, let's think through. I’m tweaking T19(x)T_{19}(x)T19​(x)'s coefficients to keep the polynomial monic and check if it simplifies into multiple irreducible components.

I'm thinking over various factorization strategies for a monic polynomial of degree 19. The goal is to ensure at least one non-linear factor and achieve at least three irreducible components in p(x)−p(y)p(x) - p(y)p(x)−p(y).

I’m updating my approach to factor x2+xy+y2−1x^2 + xy + y^2 - 1x2+xy+y2−1 over complex numbers by exploring change of variables and third roots of unity. This helps in achieving the desired polynomial structure.

I'm combining x3−xx^3 - xx3−x with x19−19xx^{19} - 19xx19−19x to tweak the polynomial's coefficients and complexity, aiming to include more non-linear factors for the desired irreducibility.

Piecing together p(x)=x19−x3−19xp(x) = x^{19} - x^3 - 19xp(x)=x19−x3−19x to ensure it’s odd, monic, and has the required linear coefficient. This targets the desired irreducible factorizations.

I’m mapping out the second factor x18+…+y18−x2−xy−y2−19x^{18} + \ldots + y^{18} - x^2 - xy - y^2 - 19x18+…+y18−x2−xy−y2−19 to see if it can split into three irreducible components.

I’m considering methods to factorize the polynomial by making it divisible by x−yx - yx−y and ensuring the remaining factor splits into exactly two irreducible components, aiming for a less complex factorization.

I’m examining the polynomial x18+x17y+…+y18x^{18} + x^{17}y + \ldots + y^{18}x18+x17y+…+y18 and noting it simplifies to x19−y19x−y\frac{x^{19} - y^{19}}{x - y}x−yx19−y19​, which might reveal further irreducible factors.

I’m mapping out the polynomial x(x18−1)−y(y18−1)x−y\frac{x(x^{18}-1) - y(y^{18}-1)}{x-y}x−yx(x18−1)−y(y18−1)​ for further factorization. This might yield just two irreducible components, which falls short of three.

I’m exploring whether the polynomial p(x)=x19−(x3)p(x) = x^19 - (x^3)p(x)=x19−(x3) can be expressed as the composition of a smaller polynomial with xkx^kxk, potentially revealing multiple irreducible components.

I'm examining if the polynomial T19(x)T_{19}(x)T19​(x) can be factored using trigonometric identities, seeking partial factorization without fully linearizing, to gain multiple irreducible components.

OK, let me see. Trigonometric identities show multiple solutions for θ\thetaθ and ϕ\phiϕ, likely leading to 19 or 38 linear factors for T19(x)T_{19}(x)T19​(x).

I’m examining how the equation arccos⁡x=±arccos⁡y+2kπ19\arccos x = \pm \arccos y + \frac{2k\pi}{19}arccosx=±arccosy+192kπ​ leads to possible multiple branches, potentially irreducible in the algebraic sense, and needing an example with at least three non-linear irreducible components.

OK, let me see. I’m pulling together details on the polynomial x3(x2−1)8+xx^3 (x^2 - 1)^8 + xx3(x2−1)8+x. It’s monic, with x19x^{19}x19 as the leading term, ensuring it meets key criteria.

I’m noting that the polynomial p(x)=x3(x2−1)8+xp(x) = x^3 (x^2 - 1)^8 + xp(x)=x3(x2−1)8+x is odd, with its linear coefficient patterning to -19.

I’m figuring out how to adjust the linear coefficient of polynomial p(x)p(x)p(x) to -19 by adding a constant factor ccc in front of xxx to ensure it’s adequately modified.

By tweaking the linear coefficient of p(x)=x3(x2−1)8+cxp(x) = x^3 (x^2 - 1)^8 + c xp(x)=x3(x2−1)8+cx to -19, it ensures the polynomial remains monic, odd, and real.

I’m working on adding a constant factor ccc to xxx in p(x)=x3(x2−1)8+cxp(x) = x^3 (x^2 - 1)^8 + c xp(x)=x3(x2−1)8+cx to ensure the linear coefficient is -19, making the polynomial monic, odd, and real.

I'm piecing together the polynomial p(x)=x3(x−1)8(x+1)8p(x) = x^3 (x-1)^8 (x+1)^8p(x)=x3(x−1)8(x+1)8 and considering how p(x)−p(y)p(x) - p(y)p(x)−p(y) might factor with factors like x−yx-yx−y.

I’m confirming p(x)p(x)p(x) is monic, odd, and real, with a linear coefficient of -19, by expanding x3(x2−1)8x^3 (x^2 -1)^8x3(x2−1)8 to set the constant ccc.

I found that for p(x) to have a coefficient of -19 for the x term, c must be -19. The polynomial, p(x), incorporates this constant and preserves the original polynomial's structure.

Below is one concrete example (over the reals) of an odd, monic, degree 19 polynomial with linear coefficient 19-19 whose zero-locus relation {p(x)=p(y)} in P1×P1\{\,p(x)=p(y)\,\}\ \text{in }\mathbb{P}^{1}\times\mathbb{P}^{1} splits into at least three irreducible components (and not all components are linear). We then compute p(19)p(19).


1. The polynomial

Consider

p(x)  =  x3(x21)8    19x.p(x) \;=\; x^{3}\,\bigl(x^{2}-1\bigr)^{8} \;-\; 19x.

We can expand

x3(x21)8  =  x19    8x17  +  28x15    56x13  +  70x11    56x9  +  28x7    8x5  +  x3.x^{3}\,\bigl(x^{2}-1\bigr)^{8} \;=\; x^{19}\;-\;8x^{17}\;+\;28x^{15}\;-\;56x^{13}\;+\;70x^{11} \;-\;56x^{9}\;+\;28x^{7}\;-\;8x^{5}\;+\;x^{3}.

Hence

p(x)  =  x19    8x17  +  28x15    56x13  +  70x11    56x9  +  28x7    8x5  +  x3    19x.p(x) \;=\; x^{19} \;-\; 8x^{17} \;+\; 28x^{15} \;-\; 56x^{13} \;+\; 70x^{11} \;-\; 56x^{9} \;+\; 28x^{7} \;-\; 8x^{5} \;+\; x^{3} \;-\; 19x.

Checking the required properties

Degree 19, monic:
The leading term is
x19x^{19}with coefficient 11.

Odd polynomial:
All exponents are odd
(19,17,,3,1),\bigl(19,\,17,\,\ldots,\,3,\,1\bigr), and the constant term is 00. Indeed,

p(x)  =  p(x),p(-x) \;=\; -\,p(x),

which makes pp an odd polynomial.

Real coefficients:
All coefficients are real integers.

Linear coefficient 19-19:
The coefficient of
xx is clearly 19-19.

At least three irreducible components, not all linear:
Over
C\mathbb{C}, the locus

X  =  {(x,y):p(x)=p(y)}    P1×P1X \;=\; \bigl\{\, (x,y) : p(x)=p(y) \,\bigr\} \;\subset\; \mathbb{P}^{1}\times\mathbb{P}^{1}

factors in part because p(x)p(x) has (for instance) the repeated factors (x21)8\bigl(x^{2}-1\bigr)^{8} in the big piece x3(x21)8x^{3}\,\bigl(x^{2}-1\bigr)^{8}. One sees that {p(x)=p(y)}\{\,p(x)=p(y)\,\} contains (at least) the divisor $${x=y\}$$ plus further factors coming from the roots x=0  or  x=±1,x = 0 \;\text{or}\; x = \pm 1, ensuring multiple irreducible components (and not all of them are hyperplanes).


2. Computing p(19)p(19)

A direct substitution x=19x=19 gives a large integer. Symbolic expansion or a computer algebra system yields

p(19)  =  1934999285524070399999639.p(19) \;=\; 1934999285524070399999639.

Hence the required value is

\boxed{\,p(19) \;=\; 1934999285524070399999639\,}. \] ​:contentReference[oaicite:0]{index=0}​

OpenAI o3 donne la bonne réponse sans faire appel à la recherche, tandis que o1 échoue à fournir une réponse correcte.

Faire progresser un raisonnement économique

Coût vs performance : o3‑mini et o4-mini
Évaluations AIME d’OpenAI o3-mini et o4-mini
Évaluations GPQA d’OpenAI o3-mini et o4-mini
Coût vs performance : o1 et o3
Évaluations AIME d’OpenAI o3 et o4
Évaluations GPQA d’OpenAI o3 et o4

OpenAI o3 et o4-mini sont nos modèles les plus intelligents jamais lancés, et ils sont aussi souvent plus efficaces que leurs prédécesseurs, OpenAI o1 et o3‑mini. Par exemple, lors de la compétition de mathématiques AIME 2025, le meilleur rapport coût/performance pour o3 est nettement supérieur à celui d’o1, et de même, le meilleur rapport pour o4-mini est nettement supérieur à celui de o3‑mini. De manière plus générale, nous anticipons que, pour la plupart des cas d’usage réels, o3 et o4-mini seront également à la fois plus intelligents et plus économiques qu’o1 et o3‑mini, respectivement.

Safety

Each improvement in model capabilities warrants commensurate improvements to safety. For OpenAI o3 and o4-mini, we completely rebuilt our safety training data, adding new refusal prompts in areas such as biological threats (biorisk), malware generation, and jailbreaks. This refreshed data has led o3 and o4-mini to achieve strong performance on our internal refusal benchmarks (e.g., instruction hierarchy, jailbreaks). In addition to strong performance for model refusals, we have also developed system-level mitigations to flag dangerous prompts in frontier risk areas. Similar to our earlier work in image generation, we trained a reasoning LLM monitor which works from human-written and interpretable safety specifications. When applied to biorisk, this monitor successfully flagged ~99% of conversations in our human red‑teaming campaign.

We stress tested both models with our most rigorous safety program to date. In accordance with our updated Preparedness Framework, we evaluated o3 and o4-mini across the three tracked capability areas covered by the Framework: biological and chemical, cybersecurity, and AI self-improvement. Based on the results of these evaluations, we have determined that both o3 and o4‑mini remain below the Framework's "High" threshold in all three categories. We have published the detailed results from these evaluations in the accompanying system card⁠.

Codex CLI : un raisonnement de pointe dans le terminal

Nous partageons également une nouvelle expérience : Codex CLI, un agent de codage léger que vous pouvez exécuter depuis votre terminal. Il fonctionne directement sur votre ordinateur et est conçu pour maximiser les capacités de raisonnement des modèles tels que o3 et o4-mini, avec un soutien à venir pour des modèles API supplémentaires comme GPT‑4.1.

Il donne accès au raisonnement multimodal des modèles depuis une ligne de commande en vous permettant de leur transférer des captures d’écran ou des esquisses basse fidélité, combiné à un accès local à votre code. Nous le considérons comme une interface minimaliste permettant de connecter nos modèles aux utilisateurs et à leurs ordinateurs. Codex CLI est entièrement en source ouverte sur github.com/openai/codex(s'ouvre dans une nouvelle fenêtre) dès aujourd'hui.

En parallèle, nous lançons une initiative dotée de 1 million de dollars visant à soutenir des projets faisant appel à Codex CLI et aux modèles OpenAI. Nous évaluerons et accepterons les demandes de subventions par tranches de 25 000 dollars américains sous forme de crédits API. Les propositions peuvent être soumettre ici.

Accès

Les utilisateurs de ChatGPT Plus, Pro et Team pourront accéder à o3, o4-mini et o4-mini-high depuis le sélecteur de modèle dès aujourd’hui. o1, o3‑mini et o3‑mini‑high ne seront quant à eux plus disponibles. Les utilisateurs de ChatGPT Enterprise et ChatGPT Edu pourront y accéder dans une semaine. Les utilisateurs gratuits peuvent essayer o4-mini en sélectionnant « Réfléchir » dans le compositeur avant d'envoyer leur requête. Les limitations applicables à l’ensemble des plans ne varient pas par rapport aux modèles précédents.

Nous prévoyons de publier OpenAI o3‑pro d’ici quelques semaines, avec une prise en charge complète des outils. D’ici là, les utilisateurs Pro conservent leur accès à o1‑pro.

Les modèles o3 et o4-mini sont également disponibles dès aujourd'hui pour les développeurs via l'API de complétions de clavardage et l'API de réponses (certains développeurs devront vérifier leur organisation(s'ouvre dans une nouvelle fenêtre) pour accéder à ces modèles). L’API de réponses prend en charge les sommaires de raisonnement, la possibilité de conserver les tokens de raisonnement autour des appels de fonction pour améliorer les performances, et prendra bientôt en charge des outils intégrés tels que la recherche Web, la recherche de fichiers et l’interpréteur de code directement dans le raisonnement du modèle. Pour commencer, explorez notre documentation(s'ouvre dans une nouvelle fenêtre) et restez à l’écoute pour d’autres mises à jour.

What's next

Today's updates reflect the direction our models are heading in: we’re converging the specialized reasoning capabilities of the o-series with more of the natural conversational abilities and tool use of the GPT‑series. By unifying these strengths, our future models will support seamless, natural conversations alongside proactive tool use and advanced problem-solving.


Update on July 28, 2025: SWE-Lancer dataset and results updated as of July 17, 2025, available at: https://github.com/openai/preparedness(s'ouvre dans une nouvelle fenêtre) and in our system cards. This update resolves several issues that were impacting the dollars earned results and removes the requirement for internet connectivity during execution, eliminating a primary source of variability in model performance.

Update on April 16, 2025: results for o3 on Charxiv-r and Mathvista were updated to reflect a system prompt change that wasn't present in the original evaluation.

Retransmission en direct

Auteur

OpenAI

Notes de bas de page

* Les chiffres d'évaluation tau-bench sont calculés en moyenne sur 5 exécutions afin de réduire la variance, et sont exécutés sans aucun outil personnalisé ni invite. Nous avons constaté que les déploiements tau-bench au détail sont plus sujets aux erreurs de modèle utilisateur. Les chiffres correspondant aux barres ombrées sont calculés avec GPT-4.1 comme modèle utilisateur, car celui-ci est nettement plus performant que GPT-4o en matière de suivi des instructions.

* SWE-bench utilise une longueur de contexte maximale de 256k, ce qui améliore le taux de résolution pour o4-mini d’environ 3 % et impacte le taux de résolution de o3 de moins de 1 %. Nous excluons également 23 échantillons qui ne peuvent pas être exécutés sur notre infrastructure interne.

** Lorsqu'on active la navigation, le modèle peut parfois trouver des réponses précises en ligne, par exemple en consultant des articles de blogue contenant des problèmes types issus d'un ensemble de données. Nous atténuons les préoccupations concernant la tricherie du modèle lors de la navigation avec deux stratégies :

  • Domaines bloqués dont nous avons observé que le modèle trichait par le passé.

  • Un modèle de raisonnement a été utilisé comme moniteur pour examiner tous les tokens de chaque tentative, afin d'identifier les comportements suspects. Un comportement suspect est défini comme « une page, un fichier ou un extrait dont l'objectif principal est de fournir la réponse exacte à cette question spécifique, par exemple, une clé de notation officielle, une fuite de “solutions” ou une discussion citant mot pour mot la réponse finale. » Le comportement bénin est défini comme « toute ressource faisant autorité qu'un être humain diligent pourrait consulter (documentation, manuels, articles savants, articles réputés), même si elle contient accidentellement la bonne réponse. » Toutes les tentatives où le moniteur a jugé le déploiement suspect sont comptées comme incorrectes. La plupart des échantillons échoués à cette vérification étaient des problèmes dont la solution exacte était disponible sur plusieurs sources Internet sans lien avec HLE.

Enfin, nos évaluations de navigation peuvent ne pas être parfaitement reproductibles dans l'API OpenAI en raison d'une différence dans les moteurs de recherche entre ChatGPT et l'API OpenAI. Ces résultats sont destinés à représenter l'expérience utilisateur de ChatGPT; cependant, en fonction de la demande, la configuration de recherche peut changer au fil du temps.

Contributeurs

Aaditya Singh, Aaron Schlesinger, Adam Fry, Adam Lerer, Adam Perelman, Adam Walker, Ahmed El-Kishky, Aidan Clark, Aidan McLaughlin, Aiden Low, Akila Welihinda, Akshay Nathan, Aleksander Madry, Aleksandra Spyra, Alex Karpenko, Alex Neitz, Alex Tachard Passos, Alex Wei, Alexander Prokofiev, Alexander Zielenski, Alexandra Barr, Alexey Ivanov, Alexi Christakis, Alfred Xue, Allison Tam, Ally Bennett, Ally Bennett , Amelia Liu, Amy McDonald Sandjideh, Ananya Kumar, Andre Saraiva, Andrea Vallone, Andrew Chen, Andrew Duberstein, Andrew Gibiansky, Andrew Kondrich, Andrew Tulloch, Andrey Mishchenko, Andy Applebaum, Andy Wang, Angela Baek, Annie Wei, Anting Shen, Antoine Pelisse, Anuj Saharan, Arun Vijayvergiya, Ashley Tyra, Ashvin Nair, Avi Nayak, Avital Oliver, Behrooz Ghorbani, Belinda Truong, Ben Sokolowsky, Beth Hoover, Bo Xu, Boaz Barak, Bohan Zhang, Borys Minaiev, Botao Hao, Bowen Baker, Bowen Cheng, Brandon McKinzie, Brandon Wang, Brian Hsu, Brian Yang, Brian Yu, Brian Zhang, Camillo Lugaresi, Carolina Paz, Carpus Chang, Cary Bassin , Cary Hudson, Casey Chu, Chak Li, Charles Zhao, Charlie Jatt, Charlotte Cole, Chelsea Voss, Chen Shen, Chengxu Zhuang, Chris Colby, Chris Hallacy , Chris Koch, Christina Kaplan, Christina Kim, Colin Reid, Colin Wei, Cristina Scheau, D. Sculley, Damien Deville, Dan Roberts, Dana Palmie, Dane Stuckey, Daniel Levine, David Hu, David Martin, David Robinson, David Sasaki, Davis Wu, Derek Chen, Dibya Bhattacharjee, Dimitris Tsipras, Dinghua Li, DJ Strouse, dmed Medina, Drew Hintz, Eddie Zhang, Edmund Wong, Elaine Ya Le, Eli Yani , Elizabeth Proehl, Emily Sokolova, Enoch Cheung, Eri Schwartz, Eric Mitchell, Eric Ning, Eric Sigler, Eric Wallace, Eugenio Panero, Evan Mays, Evgenii Nikishin, Fan Wang, Fangyuan Li, Filippo Raso, Foivos Tsimpourlas, Fouad Matin, Francis Song, Francis Zhang, Gary Yang, Gene Oden, Giambattista Parascandolo, Gildas Chabot, Grace Kim, Grace Zhao, Greg Brockman, Gregory Valiant, Guillaume Leclerc, Hadi Salman, Haitang Hu, Hannah Sheahan, Hao Sheng, Haoyu Wang, Henrique Ponde de Oliveira Pinto, Henry Aspegren, Heqing Yan, Hessam Bagherinezhad, Hongyu Ren, Hunter Lightman, Hyeonwoo Noh, Ian Kivlichan, Ian Sohl, Ignasi Clavera, Ikai Lan, Ilge Akkaya, Ilya Kostrikov, Irina Kofman, Isa Fulford, Jake Brill, Jakub Pachocki, James Betker, James Lee, James Qin, Jamie Kiros, Jason Ai, Jay Wang, Jean Harb, Jeff Mickey, Jeffrey Han, Jeffrey Wang, Jeremy Chen, Jerry Tworek, Jessica Liang, Jessica Shieh, Ji Lin, Jiahui Yu, Jianfeng Wang, Jie Tang, Jihan Yin, Jing Li, Joanne Jang, Joel Morris, Johannes Ferstad, Johannes Heidecke, John Fishbein, Jon Okun, Jonathan Gordon, Joost Huizinga, Jos Kraaijeveld, Joseph Mo, Josh Lawson , Josh Tobin, Junhua Mao, Kai Chen, Kai Hayashi, Karan Singhal, Karina Nguyen, Katy Shi, Kelly Stirman, Kenji Hata, Kenny Nguyen, Keren Gu-Lemberg, Kevin Gladstone, Kevin King, Kevin Liu, Kevin Lu, Kevin Park, Kevin Stone, Kevin Weil, Kevin Whinnery, Kevin Yu, Kote Mushegiani, Kristen Ying, Kristian Georgiev, Kshitij Gupta, Kyle Kosic, Lama Ahmad, Larry Lv, Lauren Itow, Lauren Yang, Lee Byron, Leo Chen, Leo Liu, Leon Maksin, Leyton Ho, Li Jing, Liang Xiong, Lin Yang, Linden Li, Lorenz Kuhn, Louis Feuvrier, Lu Zhang, Łukasz Kaiser, Mahmoud Eariby, Maja Trębacz, Manas Joglekar, Manoli Liodakis, Manuka Stratta, Mark Chen, Mark Hudnall, Mark Sun, Mark Wang, Martin Li, Marvin Zhang, Mateusz Litwin, Matt Jones, Matt Lim, Max Johnson, Max Schwarzer, Mayank Gupta, Meghan Shah, Mengqing Wang, Mengyuan Yan, Mia Glaese, Michael Bolin, Michael Lampe, Michael Malek, Michael Sharman, Michael Zhang, Michele Wang, Michelle Pokrass, Miguel Oom Temudo de Castro, Mihai Florian, Mike McClay, Mike Trpcic, Miki Habryn, Miles Wang, Ming Chen, Mingxuan Wang, Minnia Feng, Mitchell Gordon, Mo Bavarian, Mostafa Rohaninejad, Nacho Soto, Nakul Khanna, Nat McAleese, Natalie Staudacher, Natan LaFontaine, Neel Ajjarapu, Nick Felt, Nick Turley, Nikil Pancha, Nikita Mikhaylin, Niko Felix, Nikunj Handa, Ning Liu, Nishant Rai, Noah Jorgensen, Noam Brown, Oleg Boiko, Oleg Murk, Olivia Watkins, Olivier Godement, Oona Gleeson, Paul Ashbourne, Pavel Belov, Peter Flockhart, Peter Hoeschele, Peter Zhokhov, Philip Pronin, Phillip Guo, Phoebe Thacker, Prafulla Dhariwal, Prashanth R, Rachel Dias, Rahul Arora, Rajkumar Samuel, Rasmus Rygaard, Ravi Teja Mullapudi, Raymond Li, Raz Gaon, Reah Miyara, Reiichiro Nakano, Reimar Leike, Rennie Song, Rhythm Garg, RJ Marsan, Robert Xiong, Robin Brown, Roman Tsiupa, Rui Shu, Ruslan Nigmatullin, Saachi Jain, Saagar Patel, Sam Altman, Sam Toizer, Sam Toyer, Samir Ahmed, Samuel Miserendino, Samuel Wolrich , Sandhini Agarwal, Santiago Hernández, Sarah Dong, Savannah Heon, Scott Ethersmith, Scott Mayer McKinney, Sean Fitzgerald, Sever Banesiu, Shamez Hemani, Shengjia Zhao, Shengli Hu, Shibani Santurkar, Shreyas Krishnaswamy, Shuchao Bi, Shunyu Yao, Shuyuan Zhang, Simón Posada Fishman, Spencer Papay, Spug Golden, Srinivas Narayanan, Stanley Hsieh, Stephen Logsdon, Sundeep Tirumalareddy, Tal Stramer, Tao Wang, Tao Xin, Taylor Gordon, Tejal Patwardhan, Thibault Sottiaux, Tina Sriskandarajah, Tony Casparro, Tony Zhao, Trevor Creech, Uzair Navid Iftikhar, Valerie Qi, Vineet Kosaraju, Vishal Kuo, Vitchyr Pong, Vivek Verma, Vlad Petrov, Wenda Zhou, Wenlei Xie, Wenting Zhan, Will DePue, Will Ellsworth, William Sheu, Wyatt Thompson, Yaming Lin, Yann Dubois, Yaodong Yu, Yara Khakbaz, Yash Patil, Yifan Wu, Yilong Qin, Yining Chen, Yirui Zhang, Yo Shavit , Young Cha, Yunyun Wang, Yushi Wang, Zack Sultan, Zehao Dou, Zewei Chu, Zheng Shao, Zhigang Wang, Zhishuai Zhang, Zihao Zhang