Kind: captions
Language: en
You've probably been watching ChatGpt,
Grock, and Gemini all year, maybe even
trying to figure out which one's
actually winning this AI race.
Well, I spent months tracking every
single release, every benchmark, every
update from OpenAI, XAI, and Google Deep
Mind throughout 2025. And here's what
surprised me. There's no clear winner.
Each of these giants dominated in
completely different ways, and the
results might not be what you expect.
Welcome back to bitbiased.ai
where we do the research so you don't
have to. Join our community of AI
enthusiasts with our free weekly
newsletter. Click the link in the
description below to subscribe. You will
get the key AI news tools and learning
resources to stay ahead.
So, in this video, I'm breaking down the
entire 2025 AI landscape to show you
exactly how OpenAI's GPT series, Elon
Musk's Gro, and Google's Gemini evolved
this year.
We're going to look at the real
performance differences, the surprising
user engagement data, and what each
company's strategy actually tells us
about where AI is headed.
By the end, you'll understand which
platform dominates in which area, and
why that matters for you.
First up, let's talk about OpenAI's
massive push with not one, not two, but
three major GPT releases in just 5
months.
OpenAI's triple release strategy.
As 2025 drew to a close, OpenAI made a
move that caught everyone off guard.
Instead of the usual annual update, they
dropped three major GPT versions in
rapid succession, and each one targeted
a completely different audience.
Here's where it gets interesting.
In August, OpenAI launched GPT5,
positioning it as their smartest,
fastest, and most useful model yet. But
this wasn't just a single model release.
GPT5 came with something clever, a
unified system that includes both a fast
model for quick tasks and a deeper GPT5
thinking model for complex problems.
The magic happens with a smart router
that automatically decides which version
to use based on your query's complexity.
Think of it like having both a sprinter
and a marathon runner on your team with
a coach who knows exactly when to send
each one in.
The performance gains were substantial.
GPT5 achieved state-of-the-art results
across coding, math, writing, health,
and vision tasks. More importantly, it
dramatically reduced those frustrating
hallucinations where AI just makes
things up and improved how well it
follows your actual instructions.
Open AAI rolled this out across all chat
GPT tiers from free users getting a
limited mini version to plus and pro
subscribers accessing higher limits and
the new extended reasoning GPT5 pro
model. But wait, there's more. Just 3
months later in November, Open AAI
struck again with GPT 5.1. This wasn't
just a minor tweak. They optimized it
specifically for speed and developer
experience, making chat GPT feel faster
and more conversational
for developers working through the API.
GPT 5.1 introduced game-changing tools
like an apply patch feature and an
interactive shell specifically for
coding. This made coding assistance not
just snappier but significantly more
costefficient.
Then on November 19th, something special
happened.
Open AAI unveiled GPT 5.1 Codeex Max, a
specialized coding model that can handle
what they call project scale contexts.
We're talking millions of tokens here,
meaning it can understand and work with
entire large code bases at once. This
thing excels at multi-hour coding tasks
that would normally require constant
context refreshing. And just when you
thought they were done for the year,
December 11th brought GPT 5.2 aimed
squarely at professional knowledge work.
This next part will surprise you. On a
benchmark called GDP veil that tests
workplace tasks, GPT 5.2's thinking
model won or tied with human experts on
70.9% of tasks.
Compare that to GPT5's 38.8% 8% and
you're looking at a nearly 83%
improvement in just 4 months. It even
achieved a perfect 100% score on a 2025
competition math exam.
GPT 5.2 rolled out immediately in both
chat GPT with instant thinking and
promotes for paid users and through the
API for developers. A week later, they
dropped GPT 5.2 two codecs optimized for
agentic coding in massive code bases
handling everything from large-scale
refactors to cyber security tasks.
OpenAI positioned GPT 5.2 as their most
capable model series yet for
professional knowledge work with massive
gains in long-term reasoning tool use
and multimodal understanding.
Throughout the year, OpenAI packed chat
GPT with new capabilities that went
beyond just the core model. The
responses API now supports integrated
tools where developers can call image
generation, code interpreter, and file
search directly within prompts.
They added asynchronous background mode
for those long tasks that would normally
time out, plus chain of thought
reasoning summaries that let you peek
inside how the AI is thinking through
problems.
On the safety front, OpenAI published an
updated preparedness framework in April
and released open safety models. They're
clearly trying to balance rapid
innovation with responsible deployment.
Consumer features kept coming too. Group
chats, shopping plugins, wider app store
availability, the launch of Sora for
video generation, whisper updates, and
expanded enterprise offerings.
XAI's aggressive play with Gro. Now,
here's where things get really
interesting. While OpenAI focused on
refinement, Elon Musk's XAI took a
different approach. Build massive
compute, move fast, and integrate
everything with X.
February 18th, 2025, XAI dropped Gro 3.
This was their first major release of
the year, and it came trained on
something called the Colossus
Supercluster, which reportedly provided
10 times the compute power of previous
state-of-the-art systems. Let that sink
in for a moment. 10 times the compute.
That's not an incremental improvement.
That's a quantum leap in training
capabilities. Gro 3 focused heavily on
reasoning and knowledge, outperforming
its predecessor, Grok 2, across
benchmarks. But the standout feature was
Deep Search, a reasoningdriven web
search tool that actually thinks through
queries before searching. They rolled it
out to X Premium Plus users with an
optional think mode that exposes the
chain of thought reasoning, letting you
see exactly how Grock arrived at its
answer. XAI also launched Super Grock, a
higher tier giving access to what they
call the heavy model, their most
powerful version.
By midyear, July 9th specifically, XAI
released Grock 4. This introduced native
tool use and real-time search
integration, transforming Grock from a
chatbot into more of an interactive
agent. They offered Grock 4 heavy
through a new Super Gro Heavy
subscription for users wanting extra
power. Then in late August and
September, XAI added specialized
variants. Grock code fast one for speedy
agentic coding and Gro 4 fast for
costefficient reasoning. These let
developers access Grock through an API
with full tool integration.
The culmination came November 17th with
Grock 4.1, a refinement that
significantly improved creativity and
emotional intelligence.
And here's a telling statistic. 65% of
early users actually preferred Gro 4.1's
more coherent style over previous
versions. That's a strong user
preference signal that they nailed
something important. Grock 4.1 topped
the Elmarina leaderboard with an ELO
rating of 1483 and showed marked
reductions in factual errors and those
annoying hallucinations.
But XAI wasn't just building models.
They aggressively expanded their
platform throughout late 2025.
December 22nd brought the collections
API, basically a built-in retrieval and
rag system for developers.
December 17th saw the voice agent API
launch, letting users interact with
Grock by speech.
Earlier in the year, they'd opened a
public API beta and even released model
weights for Grock 1 back in November
2023 to invite community development.
XAI also went after specific domains.
They launched deep search for reasoning
intensive searches and in July
introduced Grock for government making
the platform available to US agencies.
This government focus became a pattern.
By late year, XAI had signed deals to
pilot AI and education with El Salvador
in December and landed national projects
with Saudi Arabia through Humane in
November.
On transparency, XAI published detailed
documentation.
In August 2025, they released a
comprehensive Gro 4 model card outlining
their risk management framework and
safety evaluations.
This matters because it shows they're
taking responsible AI seriously, not
just racing to ship features.
The funding side tells another story.
XAI raised $6 billion in series C
funding back in December 2024, fueling
this rapid expansion.
And here's where it gets fascinating.
By year end, Similar Web reported that
Grock actually led in user engagement.
Users spent approximately 8 minutes per
session on Grock compared to 6 minutes
on chat GPT. That's 33% longer
engagement time. However, Grock's
overall share of traffic remained small
at around 3% of the market. They carved
out a niche. Deep engagement with power
users, especially those who want AI
tightly integrated with live social
media data and trending topics on X.
Google DeepMind's multimodal dominance.
While OpenAI iterated and XAI built
partnerships, Google DeepMind played to
its strengths. massive infrastructure,
deep research capabilities, and
integration across every Google product
you already use.
In March 2025, Google quietly rolled out
Gemini 2.5 Pro Experimental, marketing
it as their most intelligent model, yet
with native multimodal capabilities,
chain of thought reasoning. And here's
the kicker, a 1 million token context
window.
To put that in perspective, that's like
being able to remember and reason over
entire books or massive code bases
simultaneously.
By June, Gemini 2.5 Pro and its faster
flash variant became generally available
along with flashlight for cost
efficiency. These models supported audio
output and introduced the first deep
think mode for tackling hard problems.
That summer, Google even open sourced
the Gemini CLI agent in June. letting
developers use Gemini directly from
their terminal for coding and automation
tasks.
But the real breakthrough came November
18th with Gemini 3 Pro and Gemini 3 Deep
Think. Google didn't hold back on their
claims. They touted Gemini 3 as
outperforming other AI models on 19 out
of 20 benchmarks. One particularly
striking result on humanity's last exam,
a notoriously difficult test, Gemini 3
achieved 41.0% accuracy compared to
OpenAI's GPT5 Pro at 31.6%.
That's a 30% performance advantage on
one of the hardest reasoning tests
available.
Gemini 3 topped the LM Marina rankings
upon release, showing it wasn't just
internal benchmarks. With 64K token
output and fully multimodal inputs,
handling text, images, audio, video, and
code, it enabled tasks like translating
entire long lectures or analyzing
personal videos in ways that felt
magical. The deep think mode, which
rolled out to ultra tier users, hit
unprecedented scores on tough tests,
including 45.1% on the ARC AGI2 exam, a
benchmark specifically designed to test
AGI like reasoning.
Here's what makes Google's approach
different. Integration everywhere.
By 2025, Gemini powered Google Search's
AI mode with immersive visual layouts,
Google's AI Studio, and Vert.ex XAI
platforms and even thirdparty developer
tools like cursor, GitHub, and Replit.
In November, Google launched
anti-gravity, a new Agentic IDE that
uses Gemini 3 to let AI agents
autonomously plan and code entire
applications end to end. No more
handholding through every step. Gemini
Canvas and the Gemini mobile app, which
reached over 650 million monthly users,
enabled creative workflows at massive
scale.
And remember Nano Banana?
That's Gemini 2.5 Flash image, which
went viral in August 2025 as a
photorealistic image generation model,
especially popular for 3D figure
selfies.
On the hardware side, Google
pre-integrated Gemini into devices like
Pixel phones and Samsung Galaxy devices
plus cloud services.
At IO 2024, they unified Bard and Duet
under the Gemini brand and launched an
AI premium tier, streamlining their
product lineup. But Google didn't just
build products, they pushed research
boundaries.
In July 2025, an advanced Gemini
Deepthink model achieved a gold medal
score on the International Mathematical
Olympiad, solving five out of six
problems entirely in natural language.
This wasn't about accessing calculation
tools or symbolic math engines. It
reasoned through complex math problems
the way a human would using novel
parallel reasoning and reinforcement
learning techniques.
Gemini topped numerous other benchmarks,
too.
Gemini 3 Pro scored 1487 ELO on WebDev
Arena for coding and led the vending
bench for long horizon planning.
These results showcased DeepMind's focus
on what they call agentic AI models that
can plan and execute complex multi-step
tasks autonomously.
The platform war APIs, agents, and
ecosystems.
Now, this is where the competition gets
really nuanced.
All three companies weren't just
building better models. They were
building entire ecosystems and each took
a distinctly different approach. Open
AAI's responses API integrated tools for
images, code execution, and search right
into prompts, plus asynchronous
execution for longunning tasks. They
grew their plug-in ecosystem, deepening
integration with Microsoft Copilot and
prepared to migrate from the older
assistance API to the more powerful
responses API architecture.
XAI enhanced the Grock API with
collections and voice capabilities,
enabling retrieval augmented generation
natively in the model without external
vector databases. This made Rag
workflows dramatically simpler for
developers. Google open sourced Gemini
CLI in June 2025 and offered Gemini
Canvas, AI Studio, and Vertex for
developers.
They also integrated Gemini into Chrome
and Gmail as intelligent assistants that
feel natural because they're already
embedded in tools people use daily.
These moves transform GPT, Grock, and
Gemini into agentic platforms.
Open AAI's tool using GPTs, XAI's deep
search agent, and Google's anti-gravity
coding agents all emerged in 2025,
showing that the next frontier isn't
just smarter models, but models that can
actually do work autonomously.
The 2025 wave went fully multimodal
across the board. GPT 5.2 improved image
understanding significantly.
Grock added image generation and held
multimodal chats on X, including avatar
interactions. Gemini was born
multimodal.
Gemini 3 accepts video and audio input
and can generate video and animations
through Google products. Context windows
exploded, too.
We're talking up to 1 million tokens
across all three platforms. GPT5 uses a
router plus mini model system. Gemini
handles 64K plus output with up to 1 m
context input. And Grock supports large
context for those comprehensive
reasoning tasks. On safety and
alignment, all three companies stepped
up their game in 2025. OpenAI continued
red teaming and monitoring, even
releasing custom safeguard models for
fine grain safety control. They actively
engaged with regulators, signing the EU
AI act code of practice. In July 2025,
Google DeepMind ran cooperative safety
challenges using exam benchmarks and
pledged compliance with emerging EU
rules. XAI published a riskmanagement
framework and detailed model cards like
the Gro 4 model card documenting content
filters and adversarial robustness
testing.
All three teams recruited alignment
researchers and responded to evolving
norms. For example, DeepMind sent test
results to governments as required by
the 2023 US executive orders on AI
safety. The message was clear. Move
fast, but don't break things in ways
that could cause real harm.
Strategic partnerships and market
positioning.
Here's where we start to see how
different these companies really are in
their approach to winning the AI race.
Open AAI remained tightly coupled with
Microsoft.
Azure continued as their cloud partner
and GPT technology powered Microsoft
C-Pilot, Bing Chat, and Office features
throughout 2025.
Microsoft hinted at even deeper
integration, particularly with Office
GPT tools that would make AI assistance
native to Word, Excel, and PowerPoint.
But OpenAI didn't put all their eggs in
one basket. They launched ChatgPT
enterprise education initiatives and
explored licensing deals with other
vendors.
Importantly, they navigated significant
political changes.
In early 2025, the US administration
revoked previous AI development
restrictions, favoring an
innovationfriendly approach that could
accelerate development timelines.
However, this came with signals that the
government would remove barriers and
preempt restrictive state laws, creating
a more uniform national framework.
XAI took a completely different path.
still venturebacked with their series C
funding. Elon Musk aimed to
differentiate through strategic
partnerships, especially with
governments.
The Grock platform's tight integration
with X gave it a unique social media
edge that neither OpenAI nor Google
could easily replicate.
XAI's government deals with the US,
Saudi Arabia, and El Salvador reflected
Musk's focus on broad adoption,
particularly in emerging markets where
AI infrastructure is still being built.
There's an interesting footnote here.
Musk reportedly offered nearly hundred
billion in late 2024 to acquire OpenAI's
nonprofit, underscoring his ambition to
combine OpenAI level research
capabilities with XAI's philosophy and
approach. In product terms, XAI
solidified Super Gro subscription tiers
and launched a fully public API, moving
from invitationonly access to open
availability for developers. This
democratization strategy aimed to build
a developer community quickly.
Google's strategy was all about
ubiquity. They doubled down on Gemini as
the future of their entire AI ecosystem.
Beyond consumer products, Google spun up
the AI premium subscription, which
included what was formerly called Bard
Advanced and co-branded features with
Android, Chrome, and Samsung devices. A
high-profile partnership signed in early
2024 put Gemini and Samsung Galaxy
phones, giving Google direct access to
hundreds of millions of users.
DeepMind also extended into robotics and
new domains with Gemini robotics
initiatives. By late 2025, Vertex AI,
Google's enterprise AI platform, ran
over 70% of enterprise AI workloads on
Gemini models.
Google formed partnerships with
HuggingFace and others to plug Gemini
into third party tools, making it the
default choice for many developers who
wanted something that just works with
their existing workflows.
The regulatory landscape and global
implications.
The geopolitical context shaped
everything in 2025, often in ways that
weren't immediately obvious. In January
2025, President Trump issued Executive
Order 14179
encouraging AI leadership and American
competitiveness.
Then in December 2025, he signed another
executive order creating a uniform
national AI policy while blocking
ownorous state level regulations.
These moves signaled that the US
government prioritized rapid innovation
which benefited all three companies
while pledging to protect safety,
privacy, and free speech. The FTC and
FDA also drafted guidelines on AI
content and medical claims in 2025,
directly influencing how GPT, Grock, and
Gemini could be deployed in healthcare
or advice applications.
Suddenly, making health recommendations
required meeting specific regulatory
standards.
Across the Atlantic, the EU AI act
progressed toward full implementation.
In July, the EU released its final code
of practice for general purpose AI
models, and OpenAI immediately committed
to signing it.
The act's core provisions became
effective in August 2025, imposing
strict transparency and safety
requirements on models like GPT and
Gemini. Google, OpenAI, and XAI prepared
compliance documentation, risk
assessments, and safety testing
protocols as outlined by EU guidance.
This wasn't optional. Operating in the
European market required meeting these
standards.
China presented a different challenge
entirely.
The cyerspace administration requires
generative AI providers to register and
maintain strict content control similar
to rules applied to wakesen and other
platforms. Back in 2023, this
effectively slowed any western entry of
GPT or Gemini into the Chinese market.
Though Chinese tech firms pursued their
own competing models domestically,
the regulatory winds shifted toward a
new balance.
Western governments wanted innovation
without recklessness, pushing companies
to embed safety by design rather than
bolting it on later.
On the cooperation front, all three
companies attended global AI safety
summits and contributed to new standards
on authenticity and red teaming.
Google integrated synth ID watermarking
into image outputs to combat deep fakes.
Open AAI and Deep Mind ran bug bounty
programs and partnered with security
researchers to detect potential misuse
before it happened. These efforts
directly shaped how quickly and widely
each model could be deployed in
sensitive domains like healthcare,
finance, and education going into 2026
and beyond.
Open source and community building.
The approaches to open source revealed
fundamental philosophical differences.
OpenAI's core language model weights
remained closed, but they contributed to
the research community in other ways.
In 2025, OpenAI released the GPT image
1.5 model for developers through their
API and published benchmarks like Indie
QA for Indian languages. In November,
their developer forums and hackathons
through the OpenAI fellows program grew
substantially.
OpenAI also open sourced some safety
tools like the GPTO OSS safeguard models
that let developers implement customized
safety policies. The message seemed to
be we'll keep the core models
proprietary but will give you tools to
build responsibly on top of them. XAI's
roots were partly in open source.
In 2024, they published Grock 1 model
weights and architectural details making
a statement about transparency. In 2025,
XAI released risk and safety
documentation through model cards like
the Gro 4 card. They held a limited
coding contest on X called the OpenAI
XAI challenge.
However, the Grock models themselves
remained proprietary, accessed only via
API or the grock.com interface rather
than open weights.
Still, XAI fostered community engagement
through X's platform itself.
Grock's tight coupling with X meant
trending topics and memes drove direct
user feedback, creating a unique
feedback loop.
Google embraced selective openness.
Besides Gemini CLI, DeepMind released
Gemma in February 2024, a family of
smaller Gemini derived open language
models.
While Gemini Ultra and Pro remained
closed, Google shared research freely
like their IMO benchmarking papers and
published detailed evaluation results on
safety protocols.
DeepMind safety teams ran open
competitions through the AI safety gym
and funded academic grants to encourage
external research.
Vert.ex AI provided public notebooks and
labs where the community could
experiment with Gemini without heavy
infrastructure investment.
DeepMind continued publishing in top
academic conferences, contributing to
shared knowledge on multimmodal models
and planning systems.
Their approach balanced commercial
interests with advancing the entire
field.
The real performance picture. By late
2025, the performance landscape showed
fascinating patterns that raw benchmark
numbers alone couldn't capture. OpenAI's
Chat GPT remained the most used platform
by a significant margin. ChatGpt still
had the lion's share of users and app
downloads globally.
But here's the twist. Its dominance was
slipping even as absolute usage grew.
Web traffic data showed ChatGpt's share
of visits dropped from approximately 87%
to 68% between 2015 in 2025, even though
the total number of users kept climbing.
Google's Gemini captured new users
through pure integration advantage.
Gemini's market share jumped from
roughly 5% to 18% of generative AI web
traffic over the year. The built-in
advantage of having AI inside Google
Search, Chrome, and Android gave Gemini
what analysts called a structural edge
that standalone apps simply can't match.
You don't need to download anything or
create a new account. It's just there
when you need it. XAI's Gro remained a
niche player by volume, holding
approximately 3% market share.
But it achieved something remarkable,
the highest engagement.
users spent approximately 8 minutes per
session on Grock compared to 6 minutes
on ChatgPT and similar times on Gemini.
That 33% higher engagement time
suggested that Grock users weren't just
casually trying it out. They were deeply
invested in the platform. In summary,
Chat GPT led in reach and total users.
Gemini led in convenience and
integration, and Grock led in user
engagement and session depth.
On pure technical benchmarks, Gemini 3
and GPT 5.2 traded victories depending
on the specific test. Google claimed
Gemini 3 Pro outperformed GPT5 Pro on
hard reasoning tests like humanity's
last exam, though OpenAI's GPT 5.2 led
on some agentic benchmarks, including
that perfect 100% score on a competition
math exam. Independent head-to-head
comparisons remain scarce since
companies tend to publish benchmarks
where they perform well. However, the
available evidence suggested Gemini had
a slight edge on academic reasoning
tasks, while GPT 5.2 excelled at
practical professional knowledge work.
Meanwhile, XAI's Grock held top scores
on LM arena, though it competed against
fewer challengers in that arena.
All three pushed toward low error rates
and high helpfulness in human
evaluations.
XAI reported 65% user preference for
Grock 4.1 over previous versions. Google
noted record performance on custom
benchmarks like the International
Mathematical Olympiad.
OpenAI highlighted GPT 5.2's 70.9% win
rate against human experts on workplace
tasks.
What this all means for 2026 and beyond.
By year end 2025, the AI race looked
less like a sprint with a clear winner
and more like a complex curve where
different players excelled in different
dimensions.
OpenAI's GPT 5.2 delivered what they
positioned as unmatched productivity
gains for professionals, keeping them
ahead on total user count and mainstream
adoption.
Their rapid iteration from GPT5 to 5.1
to 5.2 in just 5 months showed a company
operating at maximum velocity, willing
to release incremental improvements
quickly rather than waiting for perfect
annual releases.
Google Deep Minds Gemini gained ground
by embedding AI across platforms people
already use everyday and pushing the
frontier of reasoning and multimodal
capabilities. Their research
breakthroughs like the IMO gold medal
performance demonstrated that DeepMind's
academic roots still drove innovation.
The integration strategy meant Gemini
could grow market share simply by
existing where users already were. XAI's
Grock carved out high- growth niches in
social media integration and finance
while boasting the most intense user
engagement in the industry. The
government partnerships and emerging
market focus suggested XAI was playing a
longer game, building relationships that
could pay off as AI adoption spread
globally.
The stage is set for 2026 to bring even
bigger models, closer benchmark
competitions, and perhaps new players
entering the frey. Open source language
models continue improving rapidly.
Specialurpose AI systems for specific
domains like medicine, law, and
engineering are proliferating.
The question isn't whether AI will get
more capable. That's essentially
guaranteed.
The question is which approach wins?
OpenAI's rapid iteration and developer
focus, XAI's integration and engagement
strategy, or Google's ubiquitous
embedding across existing platforms.
One thing became crystal clear in 2025.
There's no single best AI.
There's the most widely used, the most
deeply integrated, and the most
engaging.
Depending on what you're trying to
accomplish and where you're already
spending your time, any of these three
could be the right choice. The AI race
isn't over. In many ways, it's just
beginning. And 2025 showed us that
competition drives innovation faster
than any single company could achieve
alone.
The race continues, and we're all
benefiting from it.