Claude vs GPT vs Gemini vs Open-Weight Models: The Late-2026 Update

Three months ago I published Claude vs GPT vs Gemini in 2026. I ended it by saying you should expect to revisit the decision in a few months. It took less than that.

Between September 1 and October 17, 2026, all three big labs shipped a new generation. Anthropic released four models in five weeks. OpenAI released GPT-6 in three tiers. Google skipped the Gemini 3.5 Pro launch I had described as imminent and went straight to Gemini 4. And on the open-weight side, Meta released its first major open model in a long time, while DeepSeek, Moonshot, and Alibaba kept narrowing the gap.

I also got something wrong. In July I wrote that Gemini 3.5 Pro was "about to ship any day now." It never shipped. Google rebuilt it, delayed it, and eventually released Gemini 4 Argon instead. That's a useful reminder of the rule I put in the original article's mistakes section: announced is not the same thing as available. It applies to my writing too.

This update covers what changed, what it means, and how I'd choose today. As always, treat every number here as a snapshot. Prices and benchmark scores move constantly, so check each provider's pricing page and model docs before you commit.

What Shipped Between September and October 2026

Earlier in the summer, Moonshot released the open weights for Kimi K3 (late July) and Alibaba published Qwen 3.8, including a 27B model under Apache 2.0 (mid-August). Those two matter for the open-weight section below.

The Headline: List Prices Converged

Here is the single most interesting thing about this release cycle. Put the new lineups side by side and the three labs now charge almost exactly the same at each tier.

In July the same comparison showed real spread: Sonnet 5 at $3 / $15 after its promo period, GPT-5.6 Terra at $2.50 / $15, Gemini 3.1 Pro at $2 / $12. The cheapest tier ran from Gemini Flash-Lite at $0.25 / $1.50 up to Claude Haiku 4.5 at $1 / $5. Now Haiku 5.5 and GPT-6 Luna are both at $0.10 / $0.50, a tenfold cut for Anthropic's small model, and the workhorse tier lands on the same $2 / $10 at all three labs.

This changes how you should compare models. When list prices are identical, the list price stops telling you anything. What actually decides your bill now is everything around it:

• Tokens per task, not price per token. A model that solves the problem in fewer steps, with less reasoning, and with fewer retries is cheaper even at the same rate. Anthropic's pitch for Opus 5.5 is explicitly this: close to Fable 5.1's quality while costing about 40% less to run than Opus 5, partly from lower prices and partly from using fewer tokens per step.
• Tokenizers. The same text produces different token counts on different model families, sometimes by 20 to 30 percent. Identical prices on different tokenizers are not identical prices. I explain why in why tokens matter, and you can compare token counts on your own text with the tokenizer tool on this site.
• Long-context surcharges. OpenAI bills requests over 272K input tokens at 2x input and 1.5x output for the whole request. Haiku 5.5's $0.10 / $0.50 applies up to 100K tokens and rises to $0.50 / $2.50 above that. If your workload uses long prompts, the effective price can be several times the headline.
• Caching and batch. Cached input rates and batch discounts (typically around 50% at all three labs) vary more than list prices do, and for agent loops with stable prefixes they dominate the bill. The mechanics are in prompt caching and semantic caching.
• Introductory pricing. Gemini 4 Argon's $2 / $10 is an intro rate that, according to the launch coverage, doubles to $4 / $20 on December 1, 2026. Don't build a cost model on a price with an expiry date.

The practical takeaway: price comparison now has to happen on your workload. Run a few hundred representative tasks through each candidate and compare the cost per completed task. The rate card won't tell you who's cheaper anymore.

The New Lineups

Claude (Anthropic)

The story of the Claude 5.5 family is that the middle of the lineup got much better value. Opus 5.5 costs less than Opus 5 did ($4 / $20 versus $5 / $25), and Anthropic positions it as close to Fable 5.1 on most work. For most teams, that makes Opus 5.5, not Fable, the model to reach for on hard agentic tasks, with Fable reserved for the longest and most demanding runs. Haiku 5.5's price cut is the other big change: at $0.10 / $0.50 it moves Claude from the most expensive budget tier to tied for the cheapest.

On coding, Claude's lead from the summer looks intact. Anthropic reports strong results on agentic coding suites for Opus 5.5 (SWE-bench Pro and Terminal-Bench among them), and those are vendor-reported numbers, so read them with the usual caution. Anthropic also continues to run Mythos 5.1, a restricted-access model for defensive security work, alongside the public lineup.

GPT (OpenAI)

Requests above 272K input tokens are billed at 2x input and 1.5x output across the lineup, and batch and flex processing run at about half price.

GPT-6 Astra's most meaningful improvement isn't the size of its window, which is the same as GPT-5.6 Sol's. It's how reliable it is deep inside that window. OpenAI's reported multi-needle retrieval scores between 512K and 1M tokens jumped substantially over the previous generation. If your workload really does need the model to find things in very long inputs, that's the headline.

Astra's launch also produced the best benchmark lesson of the cycle. OpenAI's launch chart showed a near-perfect score on ARC-AGI-3, but independent analysis pointed out that the number depended on a stateful adapter harness, and stateless runs scored far lower. Same model, same benchmark, wildly different numbers depending on how the test was run. Keep that in mind whenever a single benchmark score is doing all the work in a launch announcement.

The bigger change for most developers is Sol and Luna, which cut the price of OpenAI's mid and low tiers roughly in half compared with GPT-5.6.

Gemini (Google)

Google's cycle was the messiest and ended the strongest. Gemini 3.5 Pro, the flagship everyone expected in July, never shipped. Gemini 4 Argon arrived in October and, at the time of writing, sits at the top of Vals AI's composite index, one of the more independent trackers. It also looks especially strong on long-context reasoning benchmarks like GraphWalks. If you wrote off Google this summer because of the 3.5 Pro delays, it's time to re-run your evals.

Two practical notes. Gemini 2.5 Pro and 2.5 Flash were retired on October 20, 2026, so anything still pinned to them needs migrating now. And plan for Argon's price to double on December 1 when you estimate costs.

Gemini's structural advantages from July haven't changed: native multimodal input (long video, audio, huge PDFs), deep Google Workspace integration, and a very cheap Flash tier for volume.

Open-Weight Models: How Close Are They Really?

In July I mentioned open-weight models almost in passing. That was already a bit unfair, and three months later it would be a real gap in the comparison. Here are the open-weight models that matter as of late October 2026.

Three things stand out.

The gap is real but smaller than it has ever been. On Vals AI's index, the best open-weight models (DeepSeek V4.1 Flash and Kimi K3, within hundredths of a point of each other) sit roughly ten points below the top closed model at the time of writing. That's a genuine difference for the hardest agentic tasks. It's not a difference that matters for classification, extraction, summarization, routing, or most RAG answer generation, which is where most production token volume actually goes.

"Open weights" is not the same as "open source," and the license is the real decision. Of the models above, only DeepSeek V4.1 Flash (MIT) and Qwen3.8-27B (Apache 2.0) come with standard permissive licenses. Kimi K3, Qwen3.8-Max, and Muse Spark Open all use custom licenses. Muse Spark Open, for example, requires a "Built with Muse" notice on products and a separate license above a $20M monthly revenue threshold, according to coverage of the license terms. Kimi K3 and Qwen3.8-Max carry their own commercial conditions. None of these are dealbreakers for most teams, but they're legal review items, not engineering defaults. Read the actual license file before you ship.

Self-hosting the big ones is a serious infrastructure project. Kimi K3's weights are about 1.5 TB. Muse Spark Open needs on the order of 700 GB of accelerator memory depending on the format. These are models you rent from an inference provider, not models you run on a spare GPU. The genuinely self-hostable tier is the 20–30B class, where Qwen3.8-27B is currently the standout.

So when does an open-weight model beat the API? In my experience, it comes down to three cases:

1. Data can't leave your infrastructure, for regulatory, contractual, or sovereignty reasons. This is increasingly common for EU companies, and it's the case where open weights are often the only option. If you're shipping in the EU, also check what you're responsible for under the AI Act's transparency obligations, since a self-hosted stack means building provenance marking yourself.
2. Very high volume on a narrow task, where a smaller model, possibly fine-tuned or distilled, matches a frontier model's accuracy at a fraction of the cost. The decision framework is in RAG vs fine-tuning vs prompting, and the efficiency techniques in optimization of ML models.
3. You need control over the model itself: pinned versions that never change underneath you, custom fine-tunes, or access to logits and internals.

If none of those apply, the closed APIs are usually simpler and, at the new $0.10 / $0.50 budget tier, often cheaper than running your own hardware.

Benchmarks, With the Usual Grain of Salt

The ARC-AGI-3 episode is worth repeating as a general rule. A benchmark score is the output of a model plus a harness plus a scoring method. Change the harness and the same model can move by tens of points. That's also why vendor-reported numbers from different labs often aren't comparable even when they share a benchmark name: one lab reports a "strict" score, another a "partial" one, a third uses a newer version of the suite.

The only benchmark that should decide what you ship is your own, on your own tasks. The methodology is in LLM evaluation, and if your outputs are open-ended, in LLM-as-a-judge.

Which One to Use for What, Updated

Software engineering and agentic coding. Still Claude first. Opus 5.5 is now the obvious default for hard coding work, and Sonnet 5.5 covers everyday coding at half the price. Reserve Fable 5.1 for very long, very demanding agent runs. GPT-6 Astra through Codex remains the strong alternative if you want long unsupervised runs with sandboxing on by default. And if your coding agent needs to run on your own infrastructure, DeepSeek V4.1 Flash and Kimi K3 are now credible options rather than compromises.

Long documents, research, and multimodal input. Gemini 4 Argon has become the most interesting choice here, combining a flagship-class model with Gemini's long-standing multimodal and Workspace strengths. GPT-6 Astra is the alternative when long-context retrieval reliability is the deciding factor. Either way, remember that dumping a million tokens into the window is rarely the best design. Selecting the right context usually beats sending all of it, which is the whole argument of context engineering.

Everyday assistant and consumer-facing products. ChatGPT's ecosystem breadth (voice, image and video generation, the largest user base) hasn't changed. If you're building a consumer product where those features matter, GPT-6 Sol is a strong, cheaper default than it was in July.

High-volume, cost-sensitive pipelines. This is where the picture changed most. Claude Haiku 5.5 and GPT-6 Luna now both sit at $0.10 / $0.50, and Gemini's Flash tiers remain very competitive. With list prices tied, choose on accuracy per task and tokens per task, measured on your own data. Watch Haiku 5.5's 100K-token threshold if your prompts are long. For the most volume-sensitive work, also benchmark a small open-weight model like Qwen3.8-27B, or DeepSeek's hosted API with its off-peak discount.

Data residency, sovereignty, and full control. Open weights. DeepSeek V4.1 Flash (MIT) and Qwen3.8-27B (Apache 2.0) are the cleanest choices from a licensing standpoint. Kimi K3 and Muse Spark Open are strong but need a license review.

Enterprise deployments tied to a cloud. Less decisive than it used to be. The new Claude models are available on Amazon Bedrock, Google Cloud, and Microsoft Foundry, and Muse Spark Open is on all of them too. "Which cloud are we already paying for" still matters for procurement and data agreements, but it no longer locks you into a single model family.

Security-sensitive and regulated workloads. Claude remains the lab that has leaned hardest into this, including restricted security models like Mythos 5.1. For regulated data that can't leave your perimeter, a self-hosted open-weight model may be the only compliant option regardless of quality.

Strengths and Blind Spots at a Glance

Common Mistakes in This Release Cycle

• Comparing models on list price. At three of four tiers the labs now charge the same rate. Compare cost per completed task on your own workload instead.
• Budgeting on introductory pricing. Gemini 4 Argon's intro rate ends on November 30. Any cost model built on it is wrong from December.
• Ignoring pricing thresholds. OpenAI's 272K-token surcharge and Haiku 5.5's 100K-token threshold can multiply the effective price of long-prompt workloads.
• Treating "open weights" as "open source." Only some of the leading open models use permissive licenses. Read the license before you build on it.
• Trusting a single headline benchmark. The same model can swing by tens of points depending on the harness. Look for independent trackers, and then run your own evals.
• Staying pinned to retired models. Gemini 2.5 Pro and Flash were retired on October 20. Check every provider's deprecation schedule, not just Google's.
• Assuming July's conclusions still hold. Two of my own July recommendations (Gemini 3.1 Flash-Lite as the clear budget winner, Gemini 3.5 Pro as the next big flagship) are already out of date. Yours probably are too.

A Quick Decision Checklist

• [ ] Have you re-run your eval set on the September–October 2026 models instead of relying on summer results?
• [ ] Are you comparing cost per completed task, including retries, reasoning tokens, and tokenizer differences, rather than list price?
• [ ] Do any of your workloads cross a long-context pricing threshold (272K at OpenAI, 100K for Haiku 5.5)?
• [ ] Are any cost estimates based on introductory pricing that will expire?
• [ ] Is this a coding or agentic task? Start with Claude Opus 5.5 or Sonnet 5.5.
• [ ] Is this long-document or multimodal work? Evaluate Gemini 4 Argon and GPT-6 Astra.
• [ ] Is this high-volume and narrow? Compare Haiku 5.5, GPT-6 Luna, Gemini Flash, and a small open-weight model on your own data.
• [ ] Does the data have to stay on your infrastructure? Shortlist DeepSeek V4.1 Flash and Qwen3.8-27B, and send any custom-licensed model to legal review.
• [ ] Are you still calling any model that has been retired or scheduled for retirement?

Frequently Asked Questions

What are the newest models from each lab as of late October 2026?

Anthropic's current lineup is Claude Fable 5.1, Opus 5.5, Sonnet 5.5, and Haiku 5.5, all released between September 1 and October 7, 2026. OpenAI's is GPT-6 Astra (September 3) plus GPT-6 Sol and GPT-6 Luna (September 22). Google's flagship is Gemini 4 Argon, generally available since October 13, alongside the Gemini 3.8 Flash family. Among open-weight models, the most relevant are DeepSeek V4.1 Flash, Kimi K3, Qwen 3.8, and Meta's Muse Spark Open.

Which is better now, Claude, GPT, or Gemini?

None of them wins everywhere. Claude still leads on agentic coding and has the best value in the middle of its lineup. GPT-6 Astra stands out for reliability on very long inputs. Gemini 4 Argon tops one of the main independent composite indexes at the time of writing and keeps Gemini's multimodal and Workspace advantages. Pick by task, and verify on your own evaluation set.

Which model is cheapest for high-volume use?

At list price, Claude Haiku 5.5 and GPT-6 Luna are tied at $0.10 per million input tokens and $0.50 per million output tokens, with Gemini's Flash tiers also very competitive. Because list prices are now so close, the cheapest model for you depends on accuracy, tokens per task, tokenizer efficiency, caching, and pricing thresholds. Haiku 5.5's rate applies up to 100K-token prompts, for example. Measure on your own workload.

Are open-weight models good enough for production?

For many production workloads, yes. The best open-weight models sit roughly ten points behind the top closed models on independent composite indexes, which matters for the hardest agentic tasks but much less for classification, extraction, summarization, or most RAG answer generation. The bigger questions are usually the license (only some use MIT or Apache 2.0) and whether you have the infrastructure to host them.

Did Gemini 3.5 Pro ever launch?

No. After repeated delays through the summer of 2026, Google released Gemini 4 Argon as its new flagship in October instead. If you planned around Gemini 3.5 Pro, evaluate Gemini 4 Argon, and remember that its introductory price of $2 / $10 per million tokens is scheduled to rise to $4 / $20 on December 1, 2026.

How often should I re-evaluate my model choice?

At this pace, at least once per quarter, and whenever a lab you depend on ships a new generation. The cheapest way to make that sustainable is to keep a small, versioned eval set for each production use case and a model-agnostic integration layer, so switching or adding a model is a configuration change rather than a rewrite.

The Right Mental Model

In July the question was which lab to use for which task. In October the labs have matched each other on price so closely that the question has shifted. The rate card won't tell you who's cheaper anymore, the leaderboards won't tell you who's better for your task, and the gap between closed and open models is narrow enough that "can we run this ourselves?" is a real question for a lot of workloads.

What's left is the boring, reliable answer: your own tasks, your own data, your own evals, and cost measured per completed task. Build that harness once and every new release becomes an afternoon of testing instead of a week of reading launch posts. At this pace, you'll be using it again before the end of the year.

Building production AI systems? I write regularly about applied AI engineering, system architecture, and the real lessons from production deployments. Find me on LinkedIn or reach out directly at ciao@pavlo.sh.