Comparison with rivals
NegativeMany argue it loses to rivals like GLM-5.3 and GPT-6 Astra or accuse Google of bench-maxing. However, others appreciate its speed and daily practical utility compared to Claude's restrictive limits.

Gemini Flash is Google's multimodal reasoning model for code and tool workflows using documents, images, audio and video. It improves software and agent reasoning over Gemini 3.7 Flash, but higher effort can use more tokens and the minimal thinking setting is unsupported. AIB places coding in the upper group, a current comparative position rather than a guaranteed success rate. Long inputs and tool access do not guarantee complete understanding or correct actions, so verify source-grounded answers and tool results on real tasks. Compare response time and total token use with correction effort, retries and service terms on representative inputs, rather than choosing solely from a demonstration or a headline benchmark.
By effortLow 22.5Medium 30.1High 30.8
By effortLow 33.2%Medium 45.8%High 44.9%
By effortLow 80.7%Medium 84.0%High 81.3%
By effortLow 10.1%Medium 19.7%High 19.7%
By effortLow 37.1%Medium 42.1%High 47.8%
See each evaluation’s detail page for its source and last updateData corrections & removal requests
While users praise Gemini 3.8 Flash's blazing speed and competitive pricing, widespread complaints persist over severe hallucinations and unfavorable comparisons to rival coding models.
Many argue it loses to rivals like GLM-5.3 and GPT-6 Astra or accuse Google of bench-maxing. However, others appreciate its speed and daily practical utility compared to Claude's restrictive limits.
Skeptics dismiss it as benchmaxxed and underwhelming in reasoning or knowledge work. Conversely, supporters celebrate its blistering speed, low latency in multimodal live tasks, and strong agentic coding performance in tools like Antigravity.
Half the users consider it worse or no better than 3.7 Flash, noting increased token consumption and regressions. Others insist it is a clear step forward, handling complex code and prompts better than earlier versions.
Critics complain about rapid token drain and price creep compared to earlier Flash tiers. Proponents, however, consider it a high-value bargain, especially via free Google AI Studio tiers or cheap regional promotional subscriptions.
Heavily criticized for persistent hallucinations, forgetting instructions within minutes, and reckless behavior like leaking secrets or fabricating answers, with users warning against relying on it for serious projects.
I remember coding with Claude Sonnet 3.5. Then Gemini 2.5 pro was for me something like sci fi tech (and free in ai studio). Now we have Flash 3.8 that far ahead of both with insane speeds and quality outputs. Damn, last year feels like 20 years ago in this way. Cant belive that only one year is passed.
Eu estava usando o Gemini 3.7 flash com a IDE do Antigravity e já estava atendendo a 100% do que eu precisava para codar, basicamente perfeito. Agora, com esse 3.8 flash o patamar está muito mais elevado. Acredito que o refino gráfico seja uma questão de ir ajustando ao longo do desenvolvimento. Por exemplo, desenvolve a land page e depois refina os gráficos. Pelo custo x benefício, e por ser um modelo flash (rápido) a Google tá ganhando de lavada.
At this point I think I'd rather Google keep improving the Flash model than trying to compete with ChatGPT or Claude. Lately I've been using 3.8 Flash more than ChatGPT because it feels less bogged down by LLM-speech and RLHF. Gemini Flash might not be a flagship frontier model, but for all the peripheral benefits that come with a Google AI subscription, it's arguably the best value.
Já estou usando ele no gemini cli. Estou gostando e só dele me atender perfeitamente nas tarefas do dia a dia sem chegar perto do limite de 5h, eu já estou muito satisfeito pois no claude eu chegava no limite de tokens toda hora e ambos no plano de 100 reais.
A punto de pasarme al plan de $31 de kimi k , pero con estas mejoras como usar antigravity desde el celular, o personalizar agentes y la subida en calidad del 3.8 y su velocidad aun seguiré con el plan de $20 de gemini ai pro
Gemini 3.8 Flash somehow works really good for me for Assembly, Rust and Python tasks. I'm not a web dev so frontend dev doesn't concern me at all. Like I genuinely find it better than Sonnet 5 or 5.6 Terra, while being significantly faster.
It is so fast. So much faster than GPT 5.6. It is really good at brainstorming, then implementing what the brainstorming came up with. Did I say it was fast? It's fast.
Pior que eu tenho o Pro e agora usando o 3.8 é só alegria. Ele parou de ser um serviço somente de armazenamento e de fato virou uma LLM confiável A velocidade dele é a coisa mais absurda, vale muito a pena A Google é absurda, tem como não. E achei massa que eles deixaram os benchmarks que ele perdiam bem explícito, não ficaram com vergonha igual o Claude é chorão
decent model. steady and reliable workhorse, especially if you, like me, got 1.5 years Ai plan for a couple of dollars.
بصراحه اخيرا ده اول موديل ممكن تعتمد عليه في البرمجه علي مشاريع قويه عاش جوجل رجعت للمقدمه يا رجاله و اتوقع 3.9 هيبقا الصاروخ الي هرفع المنافسه علي الكل ❤🔥🔥🔥🔥
I remember coding with Claude Sonnet 3.5. Then Gemini 2.5 pro was for me something like sci fi tech (and free in ai studio). Now we have Flash 3.8 that far ahead of both with insane speeds and quality outputs. Damn, last year feels like 20 years ago in this way. Cant belive that only one year is passed.
Eu estava usando o Gemini 3.7 flash com a IDE do Antigravity e já estava atendendo a 100% do que eu precisava para codar, basicamente perfeito. Agora, com esse 3.8 flash o patamar está muito mais elevado. Acredito que o refino gráfico seja uma questão de ir ajustando ao longo do desenvolvimento. Por exemplo, desenvolve a land page e depois refina os gráficos. Pelo custo x benefício, e por ser um modelo flash (rápido) a Google tá ganhando de lavada.
At this point I think I'd rather Google keep improving the Flash model than trying to compete with ChatGPT or Claude. Lately I've been using 3.8 Flash more than ChatGPT because it feels less bogged down by LLM-speech and RLHF. Gemini Flash might not be a flagship frontier model, but for all the peripheral benefits that come with a Google AI subscription, it's arguably the best value.
Já estou usando ele no gemini cli. Estou gostando e só dele me atender perfeitamente nas tarefas do dia a dia sem chegar perto do limite de 5h, eu já estou muito satisfeito pois no claude eu chegava no limite de tokens toda hora e ambos no plano de 100 reais.
A punto de pasarme al plan de $31 de kimi k , pero con estas mejoras como usar antigravity desde el celular, o personalizar agentes y la subida en calidad del 3.8 y su velocidad aun seguiré con el plan de $20 de gemini ai pro
Gemini 3.8 Flash somehow works really good for me for Assembly, Rust and Python tasks. I'm not a web dev so frontend dev doesn't concern me at all. Like I genuinely find it better than Sonnet 5 or 5.6 Terra, while being significantly faster.
It is so fast. So much faster than GPT 5.6. It is really good at brainstorming, then implementing what the brainstorming came up with. Did I say it was fast? It's fast.
Pior que eu tenho o Pro e agora usando o 3.8 é só alegria. Ele parou de ser um serviço somente de armazenamento e de fato virou uma LLM confiável A velocidade dele é a coisa mais absurda, vale muito a pena A Google é absurda, tem como não. E achei massa que eles deixaram os benchmarks que ele perdiam bem explícito, não ficaram com vergonha igual o Claude é chorão
decent model. steady and reliable workhorse, especially if you, like me, got 1.5 years Ai plan for a couple of dollars.
بصراحه اخيرا ده اول موديل ممكن تعتمد عليه في البرمجه علي مشاريع قويه عاش جوجل رجعت للمقدمه يا رجاله و اتوقع 3.9 هيبقا الصاروخ الي هرفع المنافسه علي الكل ❤🔥🔥🔥🔥
Pretty much what I'd expect from a Gemini Flash model. More importantly, these results further confirm that 3.8 is somewhat benchmaxed, since many of its benchmark results imply it's better than Claude Opus 5 and in some cases within touching distance of Fable 5.1 and Astra, which is absolutely is not, outside of perhaps extreme edge cases.
what the hell kind of review is that? "kinda competitive" ?? Why are you even mentioning fable and sol, it obviously is inferior - and the price comparisons should be aligned with GLM-5.3-flash and GPT-5.6-Luna (it is a flash model after all) - and considering that its 5x the price of glm-5.3-flash I'm not really seeing the value here - And no mention at all about speed or token usage? I mean really token usage / price per completed task is the "value" metric you need to be using these days - From first impressions I'm giving the model the same grade as the review, "fails to meet expectations"
Gemini 3.8-flash is NOT better at coding than GLM-5.3-Flash. It messed up a simple short task (porting from c++ to a different approach also in c++) with basic errors (a bunch of messed up goto calls) and a few messages later, reintroduced the same type of errors in a different place, even though it still had the full conversation in the context window. 🤣🤣
My friend. i respect you very much. but have you tested 3.8 against 3.7 ? there is no difference! I made 130 fifferent tests all came-up identically. All had the same mistakes, all used the same amount of tokens. There is no difference! they just changed the number from 7 to 8!
Gemini is cool to daily use cases. But - for example - in website building and design still sucks, he needs 5 prompt to activate docker with wordpress when opus do it without any iteration and 3.8 still making ai slop building even with inspiration and intruction bundle. Opus doesn't have theese problems and makes very good look frontend. Both had the same skills and context. Opus was working on CC dekstop, gemini on AG (but not IDE because at the time when i was building it wasn't available.)
Estou testando o 3.8 pelo AI Studio em codificação de plataformas e o resultado é bem ruim. Entre o 3.8, 3.7, 3.6 e 3.5, até agora o melhor foi o 3.7: muito rápido e preciso nos ajustes. Em segundo, o 3.6. De resto, vou continuar testando, mas desanimado com o desempenho em meus testes...
Tenho o Gemini Pro e recebi essa atualização do 3.8, na minha opinião ele ficou bem mais burro, tá alucinando mais e só dá resposta genérica, no começo do ano eu considerava o Gemini o melhor, mas hoje estou procurando outro e vou começar a testar o Claude.
1:14 You know what…nothing against this video obviously; I’m still going to check it out for a proper review by a man of such influential status…but despite having a Pro subscription for Google AI I simply don’t give a f**k about Gemini at this point. The fact that they’ve released 3 point updates in 6 weeks and I can’t think of a single one of them that made me use Gemini is telling. I don’t know why I even have a subscription tbh 😂. All of its responses look as if it’s just responding to a Google search
Bro honestly I get hyped each time, but it feel like 3.7 flash but just slower. Gemini's context is the worst I have ever seen. If you dont say what you need with the actucal steps in detail its shit, good worker (not the best) that is very slow. shit reasoning . Luna is much better for doing work that Fable or Astra or SOL/OPUS planned.
In office tasks 3.8 flash hallucinates incredibly 😢