Anthropic pushed a new entry onto the crowded October 2026 model leaderboard on October 7, shipping Claude Haiku 5.5 at a price that undercuts its own prior small model by 90 percent for requests up to 100,000 tokens (Anthropic puts the average reduction at about 75 percent, since larger requests get a smaller discount). The release lands in the middle of a broader repricing fight across the AI industry, with Google’s Gemini 4 Argon rolling out general availability the same month and OpenAI’s GPT-5.6 Sol and GPT-6.1 Sol trading places atop live-search rankings. For developers watching token bills climb as agentic workloads scale, the Haiku 5.5 launch is less a headline model drop and more a signal of where the real competition has moved: cost per million tokens, not just raw capability.
According to Evertune’s AI model tracker, Claude Haiku 5.5 is the newest model in Anthropic’s lineup as of October 8, arriving just weeks after Claude Opus 5.5 cut its own pricing by 20 percent and stirred speculation about an eventual Opus 6. The timing is not an accident. Anthropic is pursuing a pre-IPO growth story built on run-rate revenue, and a cheap, fast model tier is the easiest way to widen the funnel of paying API customers without touching the margins on its flagship Opus line.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
Anthropic Ships Claude Haiku 5.5, Its Cheapest Model Yet
Claude Haiku 5.5 is priced at $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens, according to Anthropic’s published pricing, while prompts above 100,000 tokens cost $0.50 input and $2.50 output per million tokens. For requests up to 100,000 tokens that rate makes it 90 percent cheaper than Claude Haiku 4.5, the model it replaces at the bottom of Anthropic’s three-tier lineup of Haiku, Sonnet, and Opus, and Anthropic says the average reduction is about 75 percent because larger requests receive a smaller discount. For context, Anthropic’s current flagship, Claude Opus 5.5, sits at the opposite end of that spectrum, and the two models now bracket a pricing gap wide enough that the same API key can serve both throwaway classification tasks and frontier reasoning work at dramatically different cost profiles.
The release did not come with a flashy keynote or a named company spokesperson on record. Anthropic has increasingly treated Haiku-tier updates as quiet, almost routine drops, in contrast to the rollout fanfare around Opus releases or the Claude Opus 5.5 vs Gemini 4 Argon comparisons that dominated tech coverage earlier this quarter. That quieter approach fits the model’s purpose: Haiku is built for high-volume, latency-sensitive tasks like autocomplete, moderation, and routing, where buyers care about throughput and unit economics far more than they care about launch-day buzz.
What Changed: Pricing Breakdown for Claude Haiku 5.5
The headline number, a 90 percent price cut relative to Haiku 4.5, holds for requests up to 100,000 tokens, while Anthropic says requests above that size are 50 percent cheaper. Even so, it is the clearest signal of where Anthropic wants this model positioned. Rather than compete purely on intelligence scores, Haiku 5.5 is explicitly a cost play, aimed at developers running millions of daily calls where even fractions of a cent per request compound into real budget lines.
Input vs Output Token Costs
At $0.10 per million input tokens and $0.50 per million output tokens, Haiku 5.5’s output-to-input ratio sits at 5:1, a familiar pattern across the industry where generation costs more than ingestion. For a workload that reads a 10,000-token document and writes a 500-token summary, the math is almost trivially cheap: a fraction of a cent per call, which is precisely the use case Anthropic is targeting with high-volume customer support bots, document triage pipelines, and agentic sub-tasks spawned by larger orchestration systems.
Context Window Limits
Anthropic gives Haiku 5.5 a 1M-token context window and supports up to 128K output tokens, but the headline rate applies specifically to prompts up to 100,000 tokens, a pricing threshold that matters at a moment when rivals are pushing toward million-token context windows. Gemini 4 Argon, for instance, is powering Google’s push toward 1-million-token coding workflows in tools like Android Studio. Haiku 5.5 also adds adaptive thinking via the effort parameter, yet the pricing structure underlines that this release is meant to compete on throughput and price for shorter, high-frequency tasks.
Long-Prompt Pricing, Caching and Batch Savings
Anthropic states that Haiku 5.5 is 90% cheaper than Haiku 4.5 for requests up to 100,000 tokens and 50% cheaper for larger requests. Anthropic says about 90% of requests to the previous Haiku model were in the lower-priced category, which is why it calls Haiku 5.5 especially good value for prompts up to 100,000 tokens. For prompts above 100,000 tokens, prices are $0.50 input and $2.50 output per million tokens, and because larger requests receive a smaller discount, Anthropic says the average reduction is about 75%.
Caching and batch settings change the bill further. Haiku 5.5 cache reads cost $0.01 per million tokens for prompts up to 100,000 tokens and $0.05 above that threshold, while cache writes cost $0.125 and $0.625, respectively. Anthropic also says batch processing can save 50% and prompt caching can save up to 90%. The official Haiku 5.5 page shows Haiku 4.5 comparison figures of $1.00 per million input tokens and $5.00 per million output tokens for the displayed pricing row, but those figures should not be read as a complete Haiku 4.5 price schedule.
For teams budgeting a migration, the practical step is to measure how many of your requests stay under the 100,000-token line, since that split decides whether the 90% or the 50% discount applies. Developers can get started with claude-haiku-5-5 on the Claude Platform, and the model is also available on Amazon Web Services, Google Cloud, and Microsoft Azure.
How Claude Haiku 5.5 Performs on Current Leaderboards
Performance-wise, Haiku 5.5 is not pretending to chase the frontier. It entered one widely cited benchmark ranking in 18th place, with an Artificial Analysis Intelligence Index score of 43. That places it well behind Anthropic’s own Opus 5.5 and even behind several open-weight models released in the same week, which is consistent with how Anthropic has positioned every Haiku generation to date: fast and cheap first, capable second.
That tradeoff is the entire point. A model scoring 43 on the Artificial Analysis Intelligence Index at sub-$1-per-million-token pricing is not meant to replace Opus-tier reasoning. It is meant to replace a human intern doing data entry, or a brittle rules-based classifier, at a fraction of the latency and cost of a larger model. Judged against that bar rather than against frontier benchmarks, Haiku 5.5’s numbers look far more competitive.
The Broader October 2026 Leaderboard: Who’s on Top
Haiku 5.5 is landing into a leaderboard that has been reshuffling weekly. On BenchLeader’s overall ranking published October 15, Claude Opus 5.5 holds the top spot with a score of 71.6, narrowly ahead of Gemini 4 Argon at 71.1 and a third Anthropic model, Claude Fable 5.1, at 69.9. That three-way clustering at the top — a gap of less than three points separating first and third place — illustrates just how tight frontier competition has become, and why vendors are increasingly differentiating on price and specialization rather than chasing marginal benchmark gains that the public barely notices.
That dynamic is already visible in how Claude Opus 5.5 and Gemini 4 Argon have been pitted against each other on price rather than pure capability, and in how Claude Opus 5.5 edges out GPT-6 Astra by less than a full point on the Epoch Index. When flagship models are separated by fractions of a point, a cheap, fast tier like Haiku 5.5 becomes the more consequential product decision for most paying customers.
| Model | Tracker / Metric | Score | Rank |
|---|---|---|---|
| Claude Opus 5.5 | BenchLeader overall | 71.6 | 1st |
| Gemini 4 Argon | BenchLeader overall | 71.1 | 2nd |
| Claude Fable 5.1 | BenchLeader overall | 69.9 | 3rd |
| GPT-5.6 Sol | Arena Search Elo | 1,258 | 1st in Arena Search |
| GPT-6.1 Sol | Arena Search Elo | 1,250 | 2nd in Arena Search |
| Claude Haiku 5.5 | Artificial Analysis Intelligence Index | 43 | 18th |
| StepFun Step 5 Preview | Humanity’s Last Exam | 46.48 | Unranked (new release) |
| StepFun Step 5 Preview | AA-AnalystAgent | 35.0 | Unranked (new release) |
Gemini 4 Argon’s General Availability and Its Own Price Shift
Google’s timing compounds the pressure on Anthropic’s pricing strategy. Gemini 4 Argon became generally available to Google AI Pro and Ultra subscribers, as well as paid API customers, in October 2026. Its introductory API pricing runs at $2 per million input tokens and $10 per million output tokens through November 30, before rising to $4 and $20 per million tokens afterward. That is a steep multiple above Haiku 5.5’s rate, but Gemini 4 Argon is competing in a different weight class, aimed at the same frontier tier where Claude Opus 5.5 and GPT-6.1 Sol operate.
The scheduled November 30 price increase is itself a tell. Google appears to be using a limited-time introductory rate to win early enterprise commitments — the same playbook already visible in GPT-6.1 Sol’s own head-to-head pricing fight with Gemini 4 Argon over token limits and cost. Locking customers into integrations before a price doubles is a familiar cloud-vendor tactic, now applied to foundation model APIs.
GPT-5.6 Sol and GPT-6.1 Sol Battle for Arena Search Supremacy
OpenAI’s internal competition is just as notable as its external one. On BenchLeader’s Arena Search ranking — which evaluates AI systems that answer questions using live web search, with judges scoring both the answer and the sourcing behind it — GPT-5.6 Sol currently leads with an Elo-style score of 1,258, just ahead of the newer GPT-6.1 Sol at 1,250. That eight-point gap is close enough that it could flip within a single ranking refresh, and it raises an odd branding question: OpenAI’s newest-numbered model is not, by this particular measure, its best-performing one.
Arena Search rankings matter more than they used to because live-search grounding has become a proxy for real-world usefulness. A model that scores well on static knowledge benchmarks but fabricates sources when asked to search the web in real time is a liability for any production deployment. The fact that two OpenAI models occupy the top two spots suggests the company’s search-grounding infrastructure, not just the underlying language model, is doing a lot of the work.
A Crowded Week for New Model Releases
Haiku 5.5 is far from the only new entrant tracked during the week of October 15. Model-tracking services also logged the arrival of NVIDIA’s Nemotron 3.5 Super VL 120B A12B, Meta’s Musespark 1.3, and Liquid AI’s LFM2.5-24B-A2B. The naming conventions alone — mixing parameter counts, active-parameter notations, and version decimals — hint at how fragmented and fast-moving the open and semi-open model ecosystem has become, with vision-language, mixture-of-experts, and efficiency-tuned variants all shipping in the same seven-day window.
For builders, that pace is a double-edged sword. It means more specialized options to choose from, but it also means benchmark leaderboards are now essentially living documents, rewritten every few days as new entrants push older models down a slot. Betting an entire product roadmap on today’s leaderboard leader is a riskier proposition than it was even six months ago.
Meta’s Muse Wonder Max Arrives With an Early Benchmark Score
Meta released two new models on October 12: Muse Wonder Max and Muse Wonder Full Max. The former posted a measured BenchLeader result of 1,867.6 on the tracker’s scale, while the latter was still awaiting a published score at the time of writing. Meta’s decision to ship two closely related variants simultaneously — one scored, one pending — suggests the company is staggering its evaluation pipeline rather than holding back a finished model, a pattern increasingly common as labs race to claim a release date before a competitor’s launch window closes.
StepFun’s Step 5 Preview and the Open-Weight Licensing Debate
Chinese AI lab StepFun published weights for its Step 5 Preview model in October, but under the StepFun Model License 1.0 rather than a permissive MIT license. That distinction matters more than it might appear. A model with openly published weights is not the same as an open-source model in the license sense, and the gap between “we published the weights” and “you can freely build a commercial product on this” has become one of the more consequential fine-print issues in the current wave of Chinese open-weight releases.
On the numbers that are available, Step 5 Preview’s existing Artificial Analysis entries show 46.48 on Humanity’s Last Exam and 35.0 on AA-AnalystAgent, scores that had not changed as of the preview release and had not yet produced a corresponding row in at least one third-party model-comparison list. That lag between a model shipping and a model being fully indexed across every tracker is itself a reminder of how much manual curation still underpins leaderboards that look automated from the outside.
Why Cheap Models Matter More Than Flagship Scores Right Now
The throughline across this week’s releases is that the industry’s news cycle has shifted its center of gravity. Coverage increasingly focuses on live-search performance, model cost, coding-agent efficiency, multimodal capability, and open-weight licensing terms, rather than chasing a single permanent “best” model. Claude Haiku 5.5 is the clearest expression of that shift: it is not trying to top any leaderboard. It is trying to win the much larger, much less glamorous market of background API calls that never make a demo video but drive the bulk of production token spend.
That matters because most enterprise AI budgets are not dominated by the handful of complex reasoning queries that make headline benchmarks. They are dominated by millions of small, repetitive calls — classifying a support ticket, extracting a field from an invoice, deciding whether to escalate a chat to a human. A 90 percent price cut on exactly that workload category (Anthropic says about 90% of requests to its previous Haiku model were prompts up to 100,000 tokens) has a bigger effect on a company’s monthly AI bill than a two-point gain on a frontier benchmark most product teams will never directly test against.
Market Impact: What the Price War Means for Builders and Budgets
Anthropic’s pre-IPO positioning gives this release added weight. The company has been pitching investors on a run-rate revenue story, including figures discussed in Anthropic’s reported Q3 revenue climbing sharply ahead of a planned public listing. A cheaper Haiku tier is a volume play: more API calls at lower unit prices can still grow total revenue if adoption scales faster than the price drops, and it widens the addressable market to startups and smaller teams who found Haiku 4.5 too expensive for throwaway tasks.
Who Benefits First
The most immediate winners are teams running high-volume, latency-sensitive agent workflows — customer support triage, content moderation, structured data extraction — where a model’s raw intelligence matters less than its cost per call and response speed. Teams that have been comparing Claude’s pricing against Microsoft Copilot’s subscription model now have another data point: a sub-dollar-per-million-token API tier that undercuts most flat-rate assistant subscriptions for anyone running automated, high-frequency workloads rather than a single human typing prompts all day.
Historical Context: From Haiku 3 to Haiku 5.5
Anthropic’s Haiku line has always played the budget role in its three-tier model family, but the gap between generations has widened each cycle. Each new Haiku release has arrived cheaper and faster than its predecessor while closing some, though never all, of the capability gap with the tier above it. The 90 percent price reduction from Haiku 4.5 to Haiku 5.5 for requests up to 100,000 tokens (about 75 percent on average, by Anthropic’s own figure) is the steepest single-generation cut in that pattern to date, and it tracks a broader industry trend: as training and inference costs fall and competition intensifies, the lowest tier of every major lab’s lineup keeps getting squeezed toward near-commodity pricing, even as the top tier keeps a premium.
That squeeze is visible across the whole market, not just at Anthropic. The same week Haiku 5.5 shipped, Gemini 4 Argon’s time-limited introductory pricing and the flood of new open-weight entrants from NVIDIA, Meta, Liquid AI, and StepFun all point to a market where no vendor can hold a premium price on a mid-tier model for very long before a cheaper or more open alternative erodes it.
Competitive Comparison: Haiku 5.5 vs the Budget-Tier Field
Direct, apples-to-apples pricing comparisons across every budget-tier model released this month are difficult, since not every lab publishes rate cards the same way or at the same time. What is clear from the confirmed figures is that Claude Haiku 5.5’s $0.10 input and $0.50 output rate sits well below Gemini 4 Argon’s introductory $2 and $10 rate — though Argon is explicitly a frontier-tier model, not a budget one, so the comparison is more about market positioning than a true like-for-like matchup.
| Model | Input ($/M tokens) | Output ($/M tokens) | Positioning |
|---|---|---|---|
| Claude Haiku 5.5 | $0.10 | $0.50 | Budget tier, high-volume tasks |
| Claude Haiku 4.5 | Higher (exact rate not published in current tracker data) | Higher (exact rate not published in current tracker data) | Prior-generation budget tier |
| Gemini 4 Argon (intro, through Nov 30) | $2.00 | $10.00 | Frontier tier, promotional rate |
| Gemini 4 Argon (standard, after Nov 30) | $4.00 | $20.00 | Frontier tier, standard rate |
The missing Haiku 4.5 exact figures are a reminder that not every vendor keeps historical rate cards easy to find once a new model supersedes the old one — a minor but real transparency gap across the industry when it comes to tracking how fast prices are actually moving generation to generation.
A simple way to frame the economics: at Haiku 5.5’s published rate, processing 10 million input tokens and 2 million output tokens costs roughly $1.00 plus $1.00, or about $2.00 total. The same workload against Gemini 4 Argon’s introductory rate would cost roughly $20 plus $20, or about $40 total — a twentyfold difference that only makes sense once you remember the two models are not aimed at the same job.
input_tokens = 10_000_000
output_tokens = 2_000_000
haiku_cost = (input_tokens / 1_000_000 * 0.10) + (output_tokens / 1_000_000 * 0.50)
# haiku_cost = $2.00
argon_intro_cost = (input_tokens / 1_000_000 * 2.00) + (output_tokens / 1_000_000 * 10.00)
# argon_intro_cost = $40.00Predictions: Where the Model Cost War Goes Next
- Expect further budget-tier price cuts from OpenAI and Google within the next one to two quarters, mirroring Anthropic’s roughly 90 percent Haiku reduction, as each lab tries to own the high-volume agentic workload market rather than cede it to cheaper rivals.
- Gemini 4 Argon’s scheduled November 30 price doubling will likely draw criticism from developers who built against the introductory rate, and could push some workloads toward Haiku 5.5 or open-weight alternatives once the higher rate kicks in.
- The gap between “open weights” and genuinely open-source licensing, highlighted by StepFun’s Step 5 Preview license terms, will keep becoming a bigger factor in enterprise adoption decisions, especially for regulated industries wary of restrictive commercial-use clauses.
- Benchmark leaderboards will keep reshuffling on a near-weekly cadence as NVIDIA, Meta, Liquid AI, and Chinese labs ship new variants, making any single leaderboard snapshot increasingly a point-in-time artifact rather than a durable ranking.
- Anthropic’s pre-IPO narrative will likely keep leaning on volume metrics from cheap-tier adoption like Haiku 5.5 alongside its flagship Opus revenue figures, since investors evaluating a public listing will want evidence of a broad, not just a premium, customer base.
Frequently Asked Questions
What is Claude Haiku 5.5?
Claude Haiku 5.5 is Anthropic’s newest budget-tier language model, released October 7, 2026. It is priced at $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens (prompts above that cost $0.50 input and $2.50 output per million tokens), positioning it as a low-cost option for high-volume, latency-sensitive tasks rather than frontier reasoning work.
How much cheaper is Claude Haiku 5.5 than Haiku 4.5?
Anthropic states that Claude Haiku 5.5 is 90 percent cheaper than Claude Haiku 4.5, Anthropic’s prior budget-tier model, for requests up to 100,000 tokens and 50 percent cheaper for larger requests. Because larger requests receive a smaller discount, Anthropic puts the average reduction at about 75 percent.
How does Claude Haiku 5.5 rank on benchmarks?
Claude Haiku 5.5 entered one widely cited model ranking in 18th place, with an Artificial Analysis Intelligence Index score of 43. That places it well behind Anthropic’s own flagship Claude Opus 5.5, which topped BenchLeader’s overall ranking at 71.6 as of October 15.
Is Claude Haiku 5.5 better than Gemini 4 Argon?
They are not direct competitors in the same tier. Gemini 4 Argon is Google’s frontier-tier model, priced at $2 per million input tokens and $10 per million output tokens through an introductory period ending November 30, 2026. Claude Haiku 5.5 is a budget-tier model built for high-volume, lower-complexity tasks at a fraction of that cost.
Which AI model currently tops the leaderboard?
According to BenchLeader’s October 15 overall ranking, Claude Opus 5.5 leads with a score of 71.6, narrowly ahead of Gemini 4 Argon at 71.1 and Claude Fable 5.1 at 69.9. On BenchLeader’s separate Arena Search ranking, which measures live-search performance, GPT-5.6 Sol leads with an Elo-style score of 1,258, just ahead of GPT-6.1 Sol at 1,250.
What other AI models launched the same week as Claude Haiku 5.5?
The week of October 15, 2026 also saw the tracked release of NVIDIA’s Nemotron 3.5 Super VL 120B A12B, Meta’s Musespark 1.3 and Muse Wonder Max/Full Max, Liquid AI’s LFM2.5-24B-A2B, and StepFun’s Step 5 Preview.
Is StepFun’s Step 5 Preview open source?
StepFun published the Step 5 Preview’s weights in October 2026, but under the StepFun Model License 1.0 rather than a permissive MIT license, meaning its weights are publicly available but not released under a fully open-source license.
Why does model pricing matter more than benchmark scores right now?
Most production AI spending comes from high-volume, repetitive tasks rather than complex reasoning queries. A steep price cut on a budget-tier model, like Claude Haiku 5.5’s roughly 90 percent reduction, can lower a company’s total AI bill far more than a marginal benchmark improvement on a flagship model most workloads never actually need.
