Since GPT-4, large language models have become more varied in how they handle inputs, how they reason, and how providers package and retire them. The biggest shift is not a proven across-the-board leap in quality: it is a move toward multimodal model families, explicit fast and deeper-reasoning options, and more differentiated access. GPT-4 itself already accepted image and text inputs, so the change is an expansion of that baseline—not a transition from text-only AI to multimodal AI.
What changed after GPT-4?
Three developments define the period: providers increasingly presented models as families with different tiers; multimodal capability became a more prominent part of model design; and reasoning modes and routing became explicit product choices. Alongside those changes, model availability began to depend more visibly on whether someone was using a chat product or an API, and on the date of access.
As an Amazon Associate I earn from qualifying purchases.
These shifts make newer systems more configurable for different tasks. They do not, by themselves, establish that every newer model is better than GPT-4 for every user or workload. The cited evaluations below are provider-reported results, not a standardized comparison across companies.
How the model landscape developed
| Period | Release or change | What it shows |
|---|---|---|
| March 2023 | OpenAI announced GPT-4. Its technical report describes image and text inputs with text output. | A multimodal baseline already existed, though the documented input and output pattern was specific. |
| December 2023 | Google announced Gemini 1.0 as a multimodal model. | Multimodality became a defining part of another major model family. The announcement does not mean every model supports every modality or can generate every type of media. |
| March 2024 | Anthropic introduced Claude 3 in Haiku, Sonnet, and Opus tiers, with an initially stated 200K-token context window. | One family could offer distinct tiers rather than a single model choice. The stated context window is specific to that launch and should not be assumed for every later model or access route. |
| May 2025 onward | Anthropic documented extended thinking; its Claude 4 launch and platform notes described extended-thinking capabilities for Opus 4 and Sonnet 4. | More deliberate reasoning became an explicit capability and choice, rather than an invisible assumption about every response. |
| August 2025 | OpenAI described GPT-5 as a system combining a fast model, a deeper-reasoning model, and a real-time router. | Providers were increasingly describing a system that selects among response modes, not simply one fixed model. |
| February 13, 2026 | OpenAI said GPT-4o, GPT-4.1, GPT-4.1 mini, and o4-mini were retired from ChatGPT, while stating there were no API changes at that time. | Availability in a chat product and availability through an API can diverge. This is a dated product notice, not a claim about present access in every region or interface. |
GPT-4 was already more than a text-only model
OpenAI’s 2023 technical report describes GPT-4 as accepting image and text inputs and producing text outputs. That distinction matters: the model’s documented ability to receive images did not mean it generated images, nor does it imply support for every audio, video, or other media workflow associated with later systems.
#1 Best Overall
OpenAI also reported that GPT-4 was 82% less likely than GPT-3.5 to respond to disallowed requests and 40% more likely to produce factual responses, based on OpenAI’s internal evaluations. Its technical report said GPT-4 scored around the top 10% of test takers on a simulated bar exam. These are company-reported evaluation results; they describe particular tests and do not guarantee similar outcomes on a user’s own tasks.
What newer model families changed
More than one tier within a family
Claude 3’s Haiku, Sonnet, and Opus lineup made tier selection part of the product: a user could choose among models intended to occupy different capability and cost points. Anthropic said Opus outperformed peers on many common evaluations and reported a twofold accuracy improvement over Claude 2.1 on challenging open-ended questions. Those are Anthropic’s claims about its evaluations, not a neutral, head-to-head finding that establishes a universal ranking.
Multimodality varies by model
“Multimodal” is not a single checklist that every model satisfies. A model may accept some kinds of input without producing that kind of output, and support can differ among models in the same family or between an app and an API. Gemini 1.0’s announcement identified it as multimodal; GPT-4’s technical report specified image and text inputs with text output. For any model being considered, verify the particular version’s documented inputs and outputs rather than inferring them from its family name.
PC Slower Than It Used to Be?
A free scan shows the junk files, broken settings and background clutter dragging Windows down - then fixes them in one click.Free scan · Windows 10 & 11Crashes, No Sound, or Screen Glitches?
Random freezes, missing sound and display glitches usually trace back to one bad driver. Find and replace yours safely.Free scan · under a minuteWhy reasoning modes and routing matter
Anthropic’s documentation presents extended thinking as a way to provide enhanced reasoning for complex tasks. OpenAI’s description of GPT-5 made the choice more explicit: a fast model, a deeper-reasoning model, and a router that selects between them. These are vendor descriptions of their systems, but they point to a practical change in model design: a response may be optimized for speed or for more extensive reasoning, instead of every request following the same path.
That choice has a trade-off. A deeper-reasoning mode is intended for harder problems, while a fast mode may suit routine requests where latency matters. The material cited here does not establish a general latency or usage-cost figure, so the right comparison is the actual model and access route being used, not an assumed penalty or benefit.
How to interpret vendors’ newer performance claims
OpenAI’s August 2025 GPT-5 announcement and system card reported several comparisons. Each should be read with its baseline and test conditions attached:
- OpenAI said GPT-5 with thinking performed better than o3 on selected capabilities while using 50–80% fewer output tokens. This refers to those selected evaluations, not all tasks or all providers.
- With web search enabled, OpenAI reported that GPT-5 responses were about 45% less likely to contain a factual error than GPT-4o responses. The result is tied to the search-enabled setup and the company’s evaluation.
- OpenAI also reported that GPT-5 with thinking was about 80% less likely to contain a factual error than o3. This is an OpenAI-reported comparison, not a universal error rate.
- On a large set of production-representative conversations, OpenAI reported a decrease in deception rates for reasoning responses from 4.8% for o3 to 2.1% for GPT-5.
Numbers from different providers or announcements cannot be treated as one leaderboard when the tasks, datasets, model versions, and test conditions are not aligned. No standardized cross-vendor evaluation in the cited material establishes which model is best overall.
Recommended Free Tools
How to compare models for a real task
When choosing between models, compare the specific versions available to you on the same task and at the same time. Check these dimensions rather than relying on a launch headline:
- Inputs and outputs: Confirm whether the exact model accepts text, images, audio, video, or other inputs, and whether it can generate the output format you need.
- Reasoning option: Check whether the model offers a default response, an explicit deeper-reasoning mode, or adaptive routing. Consider whether the task warrants extra reasoning time.
- Context window: Find the documented limit for that model and access route. A larger context window does not by itself prove better performance or that all of it can be used effectively.
- Family tier: Identify whether the option is positioned for speed and lower cost, general-purpose work, or more demanding reasoning. Confirm actual cost and latency in the relevant product or API documentation.
- Availability: Check the specific app or API, region, and date. A model can be removed from one product while remaining available through another.
- Evaluation evidence: For any claimed improvement, note the publisher, model version, task or dataset, baseline, and test setup. Treat vendor comparisons as evidence about those tests, not a guarantee for unrelated work.
Are newer LLMs actually better?
They can be better suited to particular tasks because the field now offers more distinct capabilities and operating modes than the initial GPT-4 comparison suggests. But “newer” is not a result, and the available figures do not support a single all-purpose winner. The sound conclusion is narrower: model families expanded, multimodal support became more prominent, and reasoning and routing became more explicit; whether that improves a particular outcome depends on the model, task, evaluation, and access surface.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
