OpenAI ships GPT-5.6, Google delays Gemini 3.5 Pro: mapping the race in July 2026
July 2026 saw two opposite moves: OpenAI publicly launched a new model family, while Google held back its direct competitor over quality concerns. Worth looking at both together — they say a lot about how the industry is competing right now.
01OpenAI: GPT-5.6 in three sizes
On July 9, 2026, OpenAI publicly launched the GPT-5.6 family, split into three models: Sol (top-of-line), Terra (balance between cost and capability, aimed at everyday use), and Luna (the fastest and cheapest). The launch only happened after a U.S. Department of Commerce review cleared broad access to the models.
Spec sheet
Context: up to 1.05 million tokens on Sol, with up to 128 thousand output tokens
Price (API, per million input/output tokens): Sol $5 / $30 · Terra $2.5 / $15 · Luna $1 / $6
New billing twist: a price now exists for "writing" to cache, in addition to the existing 90% discount for "reading" from cache
New features: programmatic tool calls, multi-agent orchestration, and configurable cache breakpoints
The point that drew the most attention outside the spec sheet: independent benchmark reports put GPT-5.6 ahead of Claude Fable 5 on long-running agentic tasks — that is, tasks where the model needs to maintain context and make multiple chained decisions over time, not just answer a single isolated question.
02Google: the Gemini 3.5 Pro that didn't ship
Meanwhile, Google is living the opposite scenario. CEO Sundar Pichai had signaled back in May that Gemini 3.5 Pro would arrive in June 2026. The deadline passed, and as of July 21 the model simply didn't exist in Google's official catalog — the top-of-line "Pro" was still Gemini 3.1 Pro (Preview) and Gemini 2.5 Pro (stable), with only Gemini 3.5 Flash available.
The reason pointed to by internal testers: Pro was falling behind Claude Fable 5 and GPT-5.6 on advanced reasoning, code, and long-horizon tasks, plus token-efficiency issues. Google had updated the training data the previous month trying to fix this, but the result wasn't convincing — and the company chose to hold the launch rather than ship a model below expectations. The market reacted: Alphabet's stock dropped nearly 3% on the day the news went public (July 16).
03Why this matters
What the OpenAI case shows
The race has turned toward long agentic tasks, not just one-shot answer quality
Price segmentation (three sizes) has become standard practice, not a differentiator
Export/government regulation is now a real part of frontier model launch calendars
What the Google case shows
Not every giant can "respond quickly" to a competitor without sacrificing quality
Coding and long-horizon reasoning have become the benchmarks that decide reputation, not just generic metrics
A public model delay now has a direct, measurable impact on a company's market value
For anyone following these tools day to day (you, for example, using Claude in development workflows), the practical takeaway is: the gap between frontier models is narrow and shifts month to month. It's worth betting less on "model X is always better" and more on tracking, task by task, which tool actually performs best on what you need to do.
Share:
Newsletter
Learn what AI is creating for you. Don't get lost.
Get notified when a new deep dive or lesson goes up. No spam, just new posts.
Cooldecode · IA Watch · Panorama da indústria
OpenAI lança o GPT-5.6, Google atrasa o Gemini 3.5 Pro: o mapa da corrida em julho de 2026
Julho de 2026 concentrou dois movimentos opostos: a OpenAI lançou publicamente uma nova família de modelos, enquanto o Google segurou o lançamento do concorrente direto por não estar satisfeito com a qualidade. Vale entender os dois casos junto — eles dizem muito sobre como o setor está competindo agora.
01OpenAI: GPT-5.6 em três tamanhos
Em 9 de julho de 2026, a OpenAI lançou publicamente a família GPT-5.6, dividida em três modelos: Sol (topo de linha), Terra (equilíbrio entre custo e capacidade, voltado para uso do dia a dia) e Luna (o mais rápido e barato). O lançamento só ocorreu após uma revisão do Departamento de Comércio dos EUA liberar o acesso amplo aos modelos.
Ficha técnica
Contexto: até 1,05 milhão de tokens no Sol, com até 128 mil tokens de saída
Preço (API, por milhão de tokens de entrada/saída): Sol US$ 5 / US$ 30 · Terra US$ 2,5 / US$ 15 · Luna US$ 1 / US$ 6
Novidade de cobrança: passou a existir preço para "escrever" no cache, além do desconto de 90% já existente para "ler" do cache
Recursos novos: chamadas de ferramenta programáticas, orquestração multiagente e pontos de corte de cache configuráveis
O ponto que mais chamou atenção fora da ficha técnica: relatos de benchmarks independentes colocaram o GPT-5.6 à frente do Claude Fable 5 em tarefas agentivas de longa duração — ou seja, tarefas onde o modelo precisa manter contexto e tomar múltiplas decisões encadeadas ao longo do tempo, não só responder uma pergunta isolada.
02Google: o Gemini 3.5 Pro que não saiu
Enquanto isso, o Google vive o cenário oposto. O CEO Sundar Pichai havia sinalizado em maio que o Gemini 3.5 Pro chegaria em junho de 2026. O prazo passou, e até 21 de julho o modelo simplesmente não existia no catálogo oficial do Google — o topo de linha "Pro" seguia sendo o Gemini 3.1 Pro (Preview) e o Gemini 2.5 Pro (estável), com apenas o Gemini 3.5 Flash disponível.
O motivo apontado por testadores internos: o Pro estava ficando atrás do Claude Fable 5 e do GPT-5.6 em raciocínio avançado, código e tarefas de longo horizonte, além de problemas de eficiência de tokens. O Google atualizou os dados de treinamento no mês anterior tentando resolver isso, mas o resultado não convenceu — e a empresa preferiu segurar o lançamento a entregar um modelo abaixo do esperado. O mercado reagiu: as ações da Alphabet caíram quase 3% no dia em que a notícia veio a público (16 de julho).
03Por que isso importa
O que o caso OpenAI mostra
A corrida virou sobre tarefas agentivas longas, não só qualidade de resposta pontual
Segmentação por preço (três tamanhos) virou padrão de mercado, não diferencial
Regulação de exportação/governo já é parte real do calendário de lançamento de modelos de ponta
O que o caso Google mostra
Nem toda gigante consegue "responder rápido" a um concorrente sem sacrificar qualidade
Coding e raciocínio de longo horizonte viraram os benchmarks que decidem reputação, não só métricas genéricas
Atraso público de modelo já tem impacto direto e mensurável no valor de mercado da empresa
Para quem acompanha essas ferramentas no dia a dia (você, por exemplo, usando Claude em fluxos de desenvolvimento), o recado prático é: a distância entre os modelos de ponta está estreita e muda de mês a mês. Vale menos apostar fielmente em "o modelo X é sempre melhor" e mais acompanhar, tarefa por tarefa, qual ferramenta performa melhor no que você realmente precisa fazer.
Compartilhar:
Newsletter
Saiba o que a IA está criando pra você. Não fique por fora.
Seja avisado quando eu postar uma imersão ou lição nova. Sem spam, só posts novos.